- title: 'EMO: Episodic Memory Optimization for Few-Shot Meta-Learning' abstract: 'Few-shot meta-learning presents a challenge for gradient descent optimization due to the limited number of training samples per task. To address this issue, we propose an episodic memory optimization for meta-learning, we call EMO, which is inspired by the human ability to recall past learning experiences from the brain’s memory. EMO retains the gradient history of past experienced tasks in external memory, enabling few-shot learning in a memory-augmented way. By learning to retain and recall the learning process of past training tasks, EMO nudges parameter updates in the right direction, even when the gradients provided by a limited number of examples are uninformative. We prove theoretically that our algorithm converges for smooth, strongly convex objectives. EMO is generic, flexible, and model-agnostic, making it a simple plug-and-play optimizer that can be seamlessly embedded into existing optimization-based few-shot meta-learning approaches. Empirical results show that EMO scales well with most few-shot classification benchmarks and improves the performance of optimization-based meta-learning methods, resulting in accelerated convergence.' volume: 232 URL: https://proceedings.mlr.press/v232/du23a.html PDF: https://proceedings.mlr.press/v232/du23a/du23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-du23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Yingjun family: Du - given: Jiayi family: Shen - given: Xiantong family: Zhen - given: Cees G.M. family: Snoek editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 1-20 id: du23a issued: date-parts: - 2023 - 11 - 20 firstpage: 1 lastpage: 20 published: 2023-11-20 00:00:00 +0000 - title: 'Replay Buffer with Local Forgetting for Adapting to Local Environment Changes in Deep Model-Based Reinforcement Learning' abstract: 'One of the key behavioral characteristics used in neuroscience to determine whether the subject of study—be it a rodent or a human—exhibits model-based learning is effective adaptation to local changes in the environment, a particular form of adaptivity that is the focus of this work. In reinforcement learning, however, recent work has shown that modern deep model-based reinforcement-learning (MBRL) methods adapt poorly to local environment changes. An explanation for this mismatch is that MBRL methods are typically designed with sample-efficiency on a single task in mind and the requirements for effective adaptation are substantially higher, both in terms of the learned world model and the planning routine. One particularly challenging requirement is that the learned world model has to be sufficiently accurate throughout relevant parts of the state-space. This is challenging for deep-learning-based world models due to catastrophic forgetting. And while a replay buffer can mitigate the effects of catastrophic forgetting, the traditional first-in-first-out replay buffer precludes effective adaptation due to maintaining stale data. In this work, we show that a conceptually simple variation of this traditional replay buffer is able to overcome this limitation. By removing only samples from the buffer from the local neighbourhood of the newly observed samples, deep world models can be built that maintain their accuracy across the state-space, while also being able to effectively adapt to local changes in the reward function. We demonstrate this by applying our replay-buffer variation to a deep version of the classical Dyna method, as well as to recent methods such as PlaNet and DreamerV2, demonstrating that deep model-based methods can adapt effectively as well to local changes in the environment.' volume: 232 URL: https://proceedings.mlr.press/v232/rahimi-kalahroudi23a.html PDF: https://proceedings.mlr.press/v232/rahimi-kalahroudi23a/rahimi-kalahroudi23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-rahimi-kalahroudi23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Ali family: Rahimi-Kalahroudi - given: Janarthanan family: Rajendran - given: Ida family: Momennejad - given: Harm prefix: van family: Seijen - given: Sarath family: Chandar editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 21-42 id: rahimi-kalahroudi23a issued: date-parts: - 2023 - 11 - 20 firstpage: 21 lastpage: 42 published: 2023-11-20 00:00:00 +0000 - title: 'Challenging Common Assumptions about Catastrophic Forgetting and Knowledge Accumulation' abstract: 'Building learning agents that can progressively learn and accumulate knowledge is the core goal of the continual learning (CL) research field. Unfortunately, training a model on new data usually compromises the performance on past data. In the CL literature, this effect is referred to as catastrophic forgetting (CF). CF has been largely studied, and a plethora of methods have been proposed to address it on short sequences of non-overlapping tasks. In such setups, CF usually leads to a quick and significant drop in performance in past tasks. Nevertheless, despite CF, recent work showed that SGD training on linear models accumulates knowledge in a CL regression setup. This phenomenon becomes especially visible when tasks reoccur. We might then wonder if DNNs trained with SGD or any standard gradient-based optimization accumulate knowledge in such a way. Such phenomena would have interesting consequences for applying DNNs to real continual scenarios. Indeed, standard gradient-based optimization methods are significantly less computationally expensive than existing CL algorithms. In this paper, we study the progressive knowledge accumulation (KA) in DNNs trained with gradient-based algorithms in long sequences of tasks with data re-occurrence. We propose a new framework, SCoLe (Scaling Continual Learning), to investigate KA and discover that catastrophic forgetting has a limited effect on DNNs trained with SGD. When trained on long sequences with data sparsely re-occurring, the overall accuracy improves, which might be counter-intuitive given our understanding of catastrophic forgetting in CL. We empirically investigate KA in DNNs under various data occurrence frequencies. We show that the catastrophic forgetting usually observed in short scenarios does not prevent knowledge accumulation in longer ones. Moreover, propose simple and scalable strategies to increase knowledge accumulation in DNNs.' volume: 232 URL: https://proceedings.mlr.press/v232/lesort23a.html PDF: https://proceedings.mlr.press/v232/lesort23a/lesort23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-lesort23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Timothée family: Lesort - given: Oleksiy family: Ostapenko - given: Pau family: Rodríguez - given: Diganta family: Misra - given: Md Rifat family: Arefin - given: Laurent family: Charlin - given: Irina family: Rish editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 43-65 id: lesort23a issued: date-parts: - 2023 - 11 - 20 firstpage: 43 lastpage: 65 published: 2023-11-20 00:00:00 +0000 - title: 'Differentially Private Algorithms for Efficient Online Matroid Optimization' abstract: 'A matroid bandit is the online version of combinatorial optimization on a matroid, in which the learner chooses $K$ actions from a set of $L$ actions that can form a matroid basis. Many real-world applications such as recommendation systems can be modeled as matroid bandits. In such learning problems, the revealed data may involve sensitive user information. Therefore, privacy considerations are crucial. We propose two simple and practical differentially private algorithms for matroid bandits built upon the ideas of Upper Confidence Bound and Thompson Sampling. The key idea behind our first algorithm, Differentially Private Upper Confidence Bound for Matroid Bandits (DPUCB-MAT), is to construct differentially private upper confidence bounds. The second algorithm, Differentially Private Thompson Sampling for Matroid Bandits (DPTS-MAT), is based on the idea of drawing random samples from differentially private posterior distributions. Both algorithms achieve $O\left( L\ln(n)/\Delta + LK\ln(n) \min\left\{K, \ln(n) \right\}/\varepsilon \right)$ regret bounds, where $\Delta$ denotes the mean reward gap and $\varepsilon$ is the required privacy parameter. Our derived regret bounds rely on novel technical arguments that deeply explore the special structure of matroids. We show a novel way to construct ordered pairs between the played actions and the optimal actions, which contributes to decomposing a matroid bandit problem into $K$ stochastic multi-armed bandit problems. Finally, we conduct experiments to demonstrate the empirical performance of our proposed learning algorithms by using both synthetic and real-world movie-recommendation datasets.' volume: 232 URL: https://proceedings.mlr.press/v232/chandak23a.html PDF: https://proceedings.mlr.press/v232/chandak23a/chandak23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-chandak23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Kushagra family: Chandak - given: Bingshan family: Hu - given: Nidhi family: Hegde editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 66-88 id: chandak23a issued: date-parts: - 2023 - 11 - 20 firstpage: 66 lastpage: 88 published: 2023-11-20 00:00:00 +0000 - title: 'Task-Agnostic Continual Reinforcement Learning: Gaining Insights and Overcoming Challenges' abstract: 'Continual learning (CL) enables the development of models and agents that learn from a sequence of tasks while addressing the limitations of standard deep learning approaches, such as catastrophic forgetting. In this work, we investigate the factors that contribute to the performance differences between task-agnostic CL and multi-task (MTL) agents. We pose two hypotheses: (1) task-agnostic methods might provide advantages in settings with limited data, computation, or high dimensionality, and (2) faster adaptation may be particularly beneficial in continual learning settings, helping to mitigate the effects of catastrophic forgetting. To investigate these hypotheses, we introduce a replay-based recurrent reinforcement learning (3RL) methodology for task-agnostic CL agents. We assess 3RL on a synthetic task and the Meta-World benchmark, which includes 50 unique manipulation tasks. Our results demonstrate that 3RL outperforms baseline methods and can even surpass its multi-task equivalent in challenging settings with high dimensionality. We also show that the recurrent task-agnostic agent consistently outperforms or matches the performance of its transformer-based counterpart. These findings provide insights into the advantages of task-agnostic CL over task-aware MTL approaches and highlight the potential of task-agnostic methods in resource-constrained, high-dimensional, and multi-task environments.' volume: 232 URL: https://proceedings.mlr.press/v232/caccia23a.html PDF: https://proceedings.mlr.press/v232/caccia23a/caccia23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-caccia23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Massimo family: Caccia - given: Jonas family: Mueller - given: Taesup family: Kim - given: Laurent family: Charlin - given: Rasool family: Fakoor editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 89-119 id: caccia23a issued: date-parts: - 2023 - 11 - 20 firstpage: 89 lastpage: 119 published: 2023-11-20 00:00:00 +0000 - title: 'Vision-Language Models as Success Detectors' abstract: 'Detecting successful behaviour is crucial for training intelligent agents. As such, generalisable reward models are a prerequisite for agents that can learn to generalise their behaviour. In this work we focus on developing robust success detectors that leverage large, pretrained vision-language models (Flamingo, Alayrac et al. (2022)) and human reward annotations. Concretely, we treat success detection as a visual question answering (VQA) problem, denoted SuccessVQA. We study three vastly different domains: (i) interactive language-conditioned agents in a simulated household, (ii) real world robotic manipulation, and (iii) “in-the-wild” human egocentric videos. We investigate the generalisation properties of a Flamingo-based success detection model across unseen language and visual changes in the first two domains, and find that the proposed method is able to outperform bespoke reward models in out-of-distribution test scenarios with either variation. In the last domain of “in-the-wild” human videos, we show that success detection on unseen real videos presents an even more challenging generalisation task warranting future work. We hope our results encourage further work in real world success detection and reward modelling with pretrained vision-language models.' volume: 232 URL: https://proceedings.mlr.press/v232/du23b.html PDF: https://proceedings.mlr.press/v232/du23b/du23b.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-du23b.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Yuqing family: Du - given: Ksenia family: Konyushkova - given: Misha family: Denil - given: Akhil family: Raju - given: Jessica family: Landon - given: Felix family: Hill - given: Nando prefix: de family: Freitas - given: Serkan family: Cabi editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 120-136 id: du23b issued: date-parts: - 2023 - 11 - 20 firstpage: 120 lastpage: 136 published: 2023-11-20 00:00:00 +0000 - title: 'Autotelic Reinforcement Learning in Multi-Agent Environments' abstract: ' In the intrinsically motivated skills acquisition problem, the agent is set in an environment with- out any pre-defined goals and needs to acquire an open-ended repertoire of skills. To do so the agent needs to be autotelic (deriving from the Greek auto (self) and telos (end goal)): it needs to generate goals and learn to achieve them following its own intrinsic motivation rather than external supervision. Autotelic agents have so far been considered in isolation. But many applications of open-ended learning entail groups of agents. Multi-agent environments pose an additional challenge for autotelic agents: to discover and master goals that require cooperation agents must pursue them simultaneously, but they have low chances of doing so if they sample them independently. In this work, we propose a new learning paradigm for modelling such settings, the Decentralized Intrinsically Motivated Skills Acquisition Problem (Dec-IMSAP), and employ it to solve cooperative navigation tasks. First, we show that agents setting their goals independently fail to master the full diversity of goals. Then, we show that a sufficient condition for achieving this is to ensure that a group aligns its goals, i.e., the agents pursue the same cooperative goal. Our empirical analysis shows that alignment enables specialization, an efficient strategy for cooperation. Finally, we intro- duce the Goal-coordination game, a fully-decentralized emergent communication algorithm, where goal alignment emerges from the maximization of individual rewards in multi-goal cooperative environments and show that it is able to reach equal performance to a centralized training baseline that guarantees aligned goals. To our knowledge, this is the first contribution addressing the problem of intrinsically motivated multi-agent goal exploration in a decentralized training paradigm.' volume: 232 URL: https://proceedings.mlr.press/v232/nisioti23a.html PDF: https://proceedings.mlr.press/v232/nisioti23a/nisioti23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-nisioti23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Eleni family: Nisioti - given: Elias family: Masquil - given: Gautier family: Hamon - given: Clément family: Moulin-Frier editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 137-161 id: nisioti23a issued: date-parts: - 2023 - 11 - 20 firstpage: 137 lastpage: 161 published: 2023-11-20 00:00:00 +0000 - title: 'Fine-grain Inference on Out-of-Distribution Data with Hierarchical Classification' abstract: 'Machine learning methods must be trusted to make appropriate decisions in real-world environments, even when faced with out-of-distribution (OOD) samples. Many current approaches simply aim to detect OOD examples and alert the user when an unrecognized input is given. However, when the OOD sample significantly overlaps with the training data, a binary anomaly detection is not interpretable or explainable, and provides little information to the user. We propose a new model for OOD detection that makes predictions at varying levels of granularity—as the inputs become more ambiguous, the model predictions become coarser and more conservative. Consider an animal classifier that encounters an unknown bird species and a car. Both cases are OOD, but the user gains more information if the classifier recognizes that its uncertainty over the particular species is too large and predicts “bird” instead of detecting it as OOD. Furthermore, we diagnose the classifier’s performance at each level of the hierarchy improving the explainability and interpretability of the model’s predictions. We demonstrate the effectiveness of hierarchical classifiers for both fine- and coarse-grained OOD tasks.' volume: 232 URL: https://proceedings.mlr.press/v232/linderman23a.html PDF: https://proceedings.mlr.press/v232/linderman23a/linderman23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-linderman23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Randolph family: Linderman - given: Jingyang family: Zhang - given: Nathan family: Inkawhich - given: Hai family: Li - given: Yiran family: Chen editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 162-183 id: linderman23a issued: date-parts: - 2023 - 11 - 20 firstpage: 162 lastpage: 183 published: 2023-11-20 00:00:00 +0000 - title: 'The Effectiveness of World Models for Continual Reinforcement Learning' abstract: 'World models power some of the most efficient reinforcement learning algorithms. In this work, we showcase that they can be harnessed for continual learning – a situation when the agent faces changing environments. World models typically employ a replay buffer for training, which can be naturally extended to continual learning. We systematically study how different selective experience replay methods affect performance, forgetting, and transfer. We also provide recommendations regarding various modeling options for using world models. The best set of choices is called Continual-Dreamer, it is task-agnostic and utilizes the world model for continual exploration. Continual-Dreamer is sample efficient and outperforms state-of-the-art task-agnostic continual reinforcement learning methods on Minigrid and Minihack benchmarks.' volume: 232 URL: https://proceedings.mlr.press/v232/kessler23a.html PDF: https://proceedings.mlr.press/v232/kessler23a/kessler23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-kessler23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Samuel family: Kessler - given: Mateusz family: Ostaszewski - given: MichałPaweł family: Bortkiewicz - given: Mateusz family: Żarski - given: Maciej family: Wolczyk - given: Jack family: Parker-Holder - given: Stephen J. family: Roberts - given: Piotr family: Mi\loś editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 184-204 id: kessler23a issued: date-parts: - 2023 - 11 - 20 firstpage: 184 lastpage: 204 published: 2023-11-20 00:00:00 +0000 - title: 'Augmenting Autotelic Agents with Large Language Models' abstract: 'Humans learn to master open-ended repertoires of skills by imagining and practicing their own goals. This autotelic learning process, literally the pursuit of self-generated (auto) goals (telos), becomes more and more open-ended as the goals become more diverse, abstract and creative. The resulting exploration of the space of possible skills is supported by an inter-individual exploration: goal representations are culturally evolved and transmitted across individuals, in particular using language. Current artificial agents mostly rely on predefined goal representations corresponding to goal spaces that are either bounded (e.g. list of instructions), or unbounded (e.g. the space of possible visual inputs) but are rarely endowed with the ability to reshape their goal representations, to form new abstractions or to imagine creative goals. In this paper, we introduce a language model augmented autotelic agent (LMA3) that leverages a pretrained language model (LM) to support the representation, generation and learning of diverse, abstract, human-relevant goals. The LM is used as an imperfect model of human cultural transmission; an attempt to capture aspects of humans’ common-sense, intuitive physics and overall interests. Specifically, it supports three key components of the autotelic architecture: 1) a relabeler that describes the goals achieved in the agent’s trajectories, 2) a goal generator that suggests new high-level goals along with their decomposition into subgoals the agent already masters, and 3) reward functions for each of these goals. Without relying on any hand-coded goal representations, reward functions or curriculum, we show that LMA3 agents learn to master a large diversity of skills in a task-agnostic text-based environment. ' volume: 232 URL: https://proceedings.mlr.press/v232/colas23a.html PDF: https://proceedings.mlr.press/v232/colas23a/colas23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-colas23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Cédric family: Colas - given: Laetitia family: Teodorescu - given: Pierre-Yves family: Oudeyer - given: Xingdi family: Yuan - given: Marc-Alexandre family: Côté editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 205-226 id: colas23a issued: date-parts: - 2023 - 11 - 20 firstpage: 205 lastpage: 226 published: 2023-11-20 00:00:00 +0000 - title: 'Towards Single Source Domain Generalisation in Trajectory Prediction: A Motion Prior based Approach' abstract: 'Trajectory prediction is an important task in many real-world applications. However, data-driven approaches typically suffer from dramatic performance degradation when applied to unseen environments due to the inevitable domain shift brought by changes in factors such as pedestrian walking speed, the geometry of the environment, etc. In particular, when a dataset does not contain sufficient samples to determine prediction rules, the trained model can easily consider some important features as domain variant. We propose a framework that integrates a simple motion prior with deep learning to achieve, for the first time, exceptional single-source domain generalisation for trajectory prediction, in which deep learning models are only trained using a single domain and then applied to multiple novel domains. Instead of predicting the exact future positions directly from the model, we first assign a constant velocity motion prior to each pedestrian and then learn a conditional trajectory prediction model to predict residuals to the motion prior using auxiliary information from the surrounding environment. This strategy combines deep learning models with knowledge priors to simultaneously simplify training and enhance generalisation, allowing the model to focus on disentangling data-driven spatio-temporal factors while not overfitting to individual motions. We also propose a novel Train-on-Best-Motion strategy that can alleviate the adverse effects of domain shift, brought on by changes in environment, by exploiting invariances inherent to the choice of motion prior. Experiments across multiple datasets of different domains demonstrate that our approach reduces the influence of domain shift and also generalizes better to unseen environments.' volume: 232 URL: https://proceedings.mlr.press/v232/huang23a.html PDF: https://proceedings.mlr.press/v232/huang23a/huang23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-huang23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Renhao family: Huang - given: Anthony family: Tompkins - given: Maurice family: Pagnucco - given: Yang family: Song editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 227-243 id: huang23a issued: date-parts: - 2023 - 11 - 20 firstpage: 227 lastpage: 243 published: 2023-11-20 00:00:00 +0000 - title: 'RaSP: Relation-aware Semantic Prior for Weakly Supervised Incremental Segmentation' abstract: 'Class-incremental semantic image segmentation assumes multiple model updates, each enriching the model to segment new categories. This is typically carried out by providing expensive pixel-level annotations to the training algorithm for all new objects, limiting the adoption of such methods in practical applications. Approaches that solely require image-level labels offer an attractive alternative, yet, such coarse annotations lack precise information about the location and boundary of the new objects. In this paper we argue that, since classes represent not just indices but semantic entities, the conceptual relationships between them can provide valuable information that should be leveraged. We propose a weakly supervised approach that exploits such semantic relations to transfer objectness prior from the previously learned classes into the new ones, complementing the supervisory signal from image-level labels. We validate our approach on a number of continual learning tasks, and show how even a simple pairwise interaction between classes can significantly improve the segmentation mask quality of both old and new classes. We show these conclusions still hold for longer and, hence, more realistic sequences of tasks and for a challenging few-shot scenario.' volume: 232 URL: https://proceedings.mlr.press/v232/roy23a.html PDF: https://proceedings.mlr.press/v232/roy23a/roy23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-roy23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Subhankar family: Roy - given: Riccardo family: Volpi - given: Gabriela family: Csurka - given: Diane family: Larlus editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 244-269 id: roy23a issued: date-parts: - 2023 - 11 - 20 firstpage: 244 lastpage: 269 published: 2023-11-20 00:00:00 +0000 - title: 'Stabilizing Unsupervised Environment Design with a Learned Adversary' abstract: 'A key challenge in training generally-capable agents is the design of training tasks that facilitate broad generalization and robustness to environment variations. This challenge motivates the problem setting of \emph{Unsupervised Environment Design} (UED), whereby a student agent trains on an adaptive distribution of tasks proposed by a teacher agent. A pioneering approach for UED is PAIRED, which uses reinforcement learning (RL) to train a teacher policy to design tasks from scratch, making it possible to directly generate tasks that are adapted to the agent’s current capabilities. Despite its strong theoretical backing, PAIRED suffers from a variety of challenges that hinder its practical performance. Thus, state-of-the-art methods currently rely on \emph{curation} and \emph{mutation} rather than \emph{generation} of new tasks. In this work, we investigate several key shortcomings of PAIRED and propose solutions for each shortcoming. As a result, we make it possible for PAIRED to match or exceed state-of-the-art methods, producing robust agents in several challenging procedurally-generated environments, including a partially-observed maze navigation task and a continuous-control car racing environment. We believe this work motivates a renewed emphasis on UED methods based on learned models that directly generate challenging environments, potentially unlocking more open-ended RL training and, as a result, more general agents.' volume: 232 URL: https://proceedings.mlr.press/v232/mediratta23a.html PDF: https://proceedings.mlr.press/v232/mediratta23a/mediratta23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-mediratta23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Ishita family: Mediratta - given: Minqi family: Jiang - given: Jack family: Parker-Holder - given: Michael family: Dennis - given: Eugene family: Vinitsky - given: Tim family: Rocktäschel editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 270-291 id: mediratta23a issued: date-parts: - 2023 - 11 - 20 firstpage: 270 lastpage: 291 published: 2023-11-20 00:00:00 +0000 - title: 'Learning Meta Representations for Agents in Multi-Agent Reinforcement Learning' abstract: 'In multi-agent reinforcement learning, the behaviors that agents learn in a single Markov Game (MG) are typically confined to the given agent number. Every single MG induced by varying the population may possess distinct optimal joint strategies and game-specific knowledge, which are modeled independently in modern multi-agent reinforcement learning algorithms. In this work, our focus is on creating agents that can generalize across population-varying MGs. Instead of learning a unimodal policy, each agent learns a policy set comprising effective strategies across a variety of games. To achieve this, we propose Meta Representations for Agents (MRA) that explicitly models the game-common and game-specific strategic knowledge. By representing the policy sets with multi-modal latent policies, the game-common strategic knowledge and diverse strategic modes are discovered through an iterative optimization procedure. We prove that by approximately maximizing the resulting constrained mutual information objective, the policies can reach Nash Equilibrium in every evaluation MG when the latent space is sufficiently large. When deploying MRA in practical settings with limited latent space sizes, fast adaptation can be achieved by leveraging the first-order gradient information. Extensive experiments demonstrate the effectiveness of MRA in improving training performance and generalization ability in challenging evaluation games.' volume: 232 URL: https://proceedings.mlr.press/v232/zhang23a.html PDF: https://proceedings.mlr.press/v232/zhang23a/zhang23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-zhang23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Shenao family: Zhang - given: Li family: Shen - given: Lei family: Han - given: Li family: Shen editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 292-317 id: zhang23a issued: date-parts: - 2023 - 11 - 20 firstpage: 292 lastpage: 317 published: 2023-11-20 00:00:00 +0000 - title: 'Partial Hypernetworks for Continual Learning' abstract: 'Hypernetworks mitigate forgetting in continual learning (CL) by generating task-dependent weights and penalizing weight changes at a meta-model level. Unfortunately, generating all weights is not only computationally expensive for larger architectures, but also, it is not well understood whether generating all model weights is necessary. Inspired by latent replay methods in CL, we propose partial weight generation for the final layers of a model using hypernetworks while freezing the initial layers. With this objective, we first answer the question of how many layers can be frozen without compromising the final performance. Through several experiments, we empirically show that the number of layers that can be frozen is proportional to the distributional similarity in the CL stream. Then, to demonstrate the effectiveness of hypernetworks, we show that noisy streams can significantly impact the performance of latent replay methods, leading to increased forgetting when features from noisy experiences are replayed with old samples. In contrast, partial hypernetworks are more robust to noise by maintaining accuracy on previous experiences. Finally, we conduct experiments on the split CIFAR-100 and TinyImagenet benchmarks and compare different versions of partial hypernetworks to latent replay methods. We conclude that partial weight generation using hypernetworks is a promising solution to the problem of forgetting in neural networks. It can provide an effective balance between computation and final test accuracy in CL streams.' volume: 232 URL: https://proceedings.mlr.press/v232/hemati23a.html PDF: https://proceedings.mlr.press/v232/hemati23a/hemati23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-hemati23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Hamed family: Hemati - given: Vincenzo family: Lomonaco - given: Davide family: Bacciu - given: Damian family: Borth editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 318-336 id: hemati23a issued: date-parts: - 2023 - 11 - 20 firstpage: 318 lastpage: 336 published: 2023-11-20 00:00:00 +0000 - title: 'Human inductive biases for aversive continual learning — a hierarchical Bayesian nonparametric model' abstract: 'Humans and animals often display remarkable continual learning abilities, adapting quickly to changing environments while retaining, reusing, and accumulating old knowledge over a lifetime. Unfortunately, in environments with adverse outcomes, the inductive biases supporting such forms of learning can turn maladaptive, yielding persistent negative beliefs that are hard to extinguish, such as those prevalent in anxiety disorders. Here, we present and model human behavioral data from a fear-conditioning task with changing latent contexts, in which participants had to predict whether visual stimuli would be followed by an aversive scream. We show that participants’ learning in our task spans three different regimes — with old knowledge either being updated, discarded (forgotten) or retained and reused in new contexts (remembered) by different participants. The latter regime corresponds to (maladaptive) spontaneous recovery of fear. We demonstrate using simulations that these behavioral regimes can be captured by varying inductive biases in Bayesian non-parametric models of contextual learning. In particular, we show that the “remembering" regime can be produced by “persistent" variants of hierarchical Dirichlet process priors over contexts and negatively biased “deterministic" beta distribution priors over outcomes. Such inductive biases correspond well to widely observed “core beliefs" that may have adaptive value in some lifelong-learning environments, at the cost of being maladaptive in other environments and tasks such as ours. Our work offers a tractable window into human inductive biases for continual learning algorithms, and could potentially help identify individual differences in learning strategies relevant for response to psychotherapy.' volume: 232 URL: https://proceedings.mlr.press/v232/pisupati23a.html PDF: https://proceedings.mlr.press/v232/pisupati23a/pisupati23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-pisupati23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Sashank family: Pisupati - given: Isabel M family: Berwian - given: Jamie family: Chiu - given: Yongjing family: Ren - given: Yael family: Niv editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 337-346 id: pisupati23a issued: date-parts: - 2023 - 11 - 20 firstpage: 337 lastpage: 346 published: 2023-11-20 00:00:00 +0000 - title: 'Prospective Learning: Principled Extrapolation to the Future' abstract: 'Learning is a process which can update decision rules, based on past experience, such that future performance improves. Traditionally, machine learning is often evaluated under the assumption that the future will be identical to the past in distribution or change adversarially. But these assumptions can be either too optimistic or pessimistic for many problems in the real world. Real world scenarios evolve over multiple spatiotemporal scales with partially predictable dynamics. Here we reformulate the learning problem to one that centers around this idea of dynamic futures that are partially learnable. We conjecture that certain sequences of tasks are not retrospectively learnable (in which the data distribution is fixed), but are prospectively learnable (in which distributions may be dynamic), suggesting that prospective learning is more difficult in kind than retrospective learning. We argue that prospective learning more accurately characterizes many real world problems that (1) currently stymie existing artificial intelligence solutions and/or (2) lack adequate explanations for how natural intelligences solve them. Thus, studying prospective learning will lead to deeper insights and solutions to currently vexing challenges in both natural and artificial intelligences.' volume: 232 URL: https://proceedings.mlr.press/v232/de-silva23a.html PDF: https://proceedings.mlr.press/v232/de-silva23a/de-silva23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-de-silva23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Ashwin family: De Silva - given: Rahul family: Ramesh - given: Lyle family: Ungar - given: Marshall Hussain family: Shuler - given: Noah J. family: Cowan - given: Michael family: Platt - given: Chen family: Li - given: Leyla family: Isik - given: Seung-Eon family: Roh - given: Adam family: Charles - given: Archana family: Venkataraman - given: Brian family: Caffo - given: Javier J. family: How - given: Justus M family: Kebschull - given: John W. family: Krakauer - given: Maxim family: Bichuch - given: Kaleab Alemayehu family: Kinfu - given: Eva family: Yezerets - given: Dinesh family: Jayaraman - given: Jong M. family: Shin - given: Soledad family: Villar - given: Ian family: Phillips - given: Carey E. family: Priebe - given: Thomas family: Hartung - given: Michael I. family: Miller - given: Jayanta family: Dey - given: Ningyuan family: Huang - given: Eric family: Eaton - given: Ralph family: Etienne-Cummings - given: Elizabeth L. family: Ogburn - given: Randal family: Burns - given: Onyema family: Osuagwu - given: Brett family: Mensh - given: Alysson R. family: Muotri - given: Julia family: Brown - given: Chris family: White - given: Weiwei family: Yang - given: Andrei A. Rusu Timothy family: Verstynen - given: Konrad P. family: Kording - given: Pratik family: Chaudhari - given: Joshua T. family: Vogelstein editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 347-357 id: de-silva23a issued: date-parts: - 2023 - 11 - 20 firstpage: 347 lastpage: 357 published: 2023-11-20 00:00:00 +0000 - title: 'Embodied Active Learning of Relational State Abstractions for Bilevel Planning' abstract: 'State abstraction is an effective technique for planning in robotics environments with continuous states and actions, long task horizons, and sparse feedback. In object-oriented environments, predicates are a particularly useful form of state abstraction because of their compatibility with symbolic planners and their capacity for relational generalization. However, to plan with predicates, the agent must be able to interpret them in continuous environment states (i.e., ground the symbols). Manually programming predicate interpretations can be difficult, so we would instead like to learn them from data. We propose an embodied active learning paradigm where the agent learns predicate interpretations through online interaction with an expert. For example, after taking actions in a block stacking environment, the agent may ask the expert: "Is On(block1, block2) true?” From this experience, the agent learns to plan: it learns neural predicate interpretations, symbolic planning operators, and neural samplers that can be used for bilevel planning. During exploration, the agent plans to learn: it uses its current models to select actions towards generating informative expert queries. We learn predicate interpretations as ensembles of neural networks and use their entropy to measure the informativeness of potential queries. We evaluate this approach in three robotic environments and find that it consistently outperforms six baselines while exhibiting sample efficiency in two key metrics: number of environment interactions, and number of queries to the expert.' volume: 232 URL: https://proceedings.mlr.press/v232/li23a.html PDF: https://proceedings.mlr.press/v232/li23a/li23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-li23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Amber family: Li - given: Tom family: Silver editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 358-375 id: li23a issued: date-parts: - 2023 - 11 - 20 firstpage: 358 lastpage: 375 published: 2023-11-20 00:00:00 +0000 - title: 'Dealing With Non-stationarity in Decentralized Cooperative Multi-Agent Deep Reinforcement Learning via Multi-Timescale Learning' abstract: 'Decentralized cooperative multi-agent deep reinforcement learning (MARL) can be a versatile learning framework, particularly in scenarios where centralized training is either not possible or not practical. One of the critical challenges in decentralized deep MARL is the non-stationarity of the learning environment when multiple agents are learning concurrently. A commonly used and efficient scheme for decentralized MARL is independent learning in which agents concurrently update their policies independently of each other. We first show that independent learning does not always converge, while sequential learning where agents update their policies one after another in a sequence is guaranteed to converge to an agent-by-agent optimal solution. In sequential learning, when one agent updates its policy, all other agent’s policies are kept fixed, alleviating the challenge of non-stationarity due to simultaneous updates in other agents’ policies. However, it can be slow because only one agent is learning at any time. Therefore it might also not always be practical. In this work, we propose a decentralized cooperative MARL algorithm based on multi-timescale learning. In multi-timescale learning, all agents learn simultaneously, but at different learning rates. In our proposed method, when one agent updates its policy, other agents are allowed to update their policies as well, but at a slower rate. This speeds up sequential learning, while also minimizing non-stationarity caused by other agents updating concurrently. Multi-timescale learning outperforms state-of-the-art decentralized learning methods on a set of challenging multi-agent cooperative tasks in the epymarl benchmark. This can be seen as a first step towards more general decentralized cooperative deep MARL methods based on multi-timescale learning.' volume: 232 URL: https://proceedings.mlr.press/v232/nekoei23a.html PDF: https://proceedings.mlr.press/v232/nekoei23a/nekoei23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-nekoei23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Hadi family: Nekoei - given: Akilesh family: Badrinaaraayanan - given: Amit family: Sinha - given: Mohammad family: Amini - given: Janarthanan family: Rajendran - given: Aditya family: Mahajan - given: Sarath family: Chandar editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 376-398 id: nekoei23a issued: date-parts: - 2023 - 11 - 20 firstpage: 376 lastpage: 398 published: 2023-11-20 00:00:00 +0000 - title: 'PlaStIL: Plastic and Stable Exemplar-Free Class-Incremental Learning' abstract: 'Plasticity and stability are needed in class-incremental learning in order to learn from new data while preserving past knowledge. Due to catastrophic forgetting, finding a compromise between these two properties is particularly challenging when no memory buffer is available. Mainstream methods need to store two deep models since they integrate new classes using fine tuning with knowledge distillation from the previous incremental state. We propose a method which has similar number of parameters but distributes them differently in order to find a better balance between plasticity and stability. Following an approach already deployed by transfer-based incremental methods, we freeze the feature extractor after the initial state. Classes in the oldest incremental states are trained with this frozen extractor to ensure stability. Recent classes are predicted using partially fine-tuned models in order to introduce plasticity. Our proposed plasticity layer can be incorporated to any transfer-based method designed for exemplar-free incremental learning, and we apply it to two such methods. Evaluation is done with three large-scale datasets. Results show that performance gains are obtained in all tested configurations compared to existing methods. ' volume: 232 URL: https://proceedings.mlr.press/v232/petit23a.html PDF: https://proceedings.mlr.press/v232/petit23a/petit23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-petit23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Grégoire family: Petit - given: Adrian family: Popescu - given: Eden family: Belouadah - given: David family: Picard - given: Bertrand family: Delezoide editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 399-414 id: petit23a issued: date-parts: - 2023 - 11 - 20 firstpage: 399 lastpage: 414 published: 2023-11-20 00:00:00 +0000 - title: 'Partial Index Tracking: A Meta-Learning Approach' abstract: 'Partial index tracking aims to cost effectively replicate the performance of a benchmark index by using a small number of assets. It is usually formulated as a regression problem, but solving it subject to real-world constraints is non-trivial. For example, the common $\ell_1$ regularised model for sparse regression (i.e., LASSO) is not compatible with those constraints. In this work, we meta-learn a sparse asset selection and weighting strategy that subsequently enables effective partial index tracking by quadratic programming. In particular, we adopt an element-wise $\ell_1$ norm for sparse regularisation, and meta-learn the weight for each $\ell_1$ term. Rather than meta-learning a fixed set of hyper-parameters, we meta-learn an inductive predictor for them based on market history, which allows generalisation over time, and even across markets. Experiments are conducted on four indices from different countries, and the empirical results demonstrate the superiority of our method over other baselines. The code is released at https://github.com/qmfin/MetaIndexTracker.' volume: 232 URL: https://proceedings.mlr.press/v232/yang23a.html PDF: https://proceedings.mlr.press/v232/yang23a/yang23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-yang23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Yongxin family: Yang - given: Timothy family: Hospedales editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 415-436 id: yang23a issued: date-parts: - 2023 - 11 - 20 firstpage: 415 lastpage: 436 published: 2023-11-20 00:00:00 +0000 - title: 'Class-Incremental Learning with Repetition' abstract: 'Real-world data streams naturally include the repetition of previous concepts. From a Continual Learning (CL) perspective, repetition is a property of the environment and, unlike replay, cannot be controlled by the agent. Nowadays, the Class-Incremental (CI) scenario represents the leading test-bed for assessing and comparing CL strategies. This scenario type is very easy to use, but it never allows revisiting previously seen classes, thus completely neglecting the role of repetition. We focus on the family of Class-Incremental with Repetition (CIR) scenario, where repetition is embedded in the definition of the stream. We propose two stochastic stream generators that produce a wide range of CIR streams starting from a single dataset and a few interpretable control parameters. We conduct the first comprehensive evaluation of repetition in CL by studying the behavior of existing CL strategies under different CIR streams. We then present a novel replay strategy that exploits repetition and counteracts the natural imbalance present in the stream. On both CIFAR100 and TinyImageNet, our strategy outperforms other replay approaches, which are not designed for environments with repetition.' volume: 232 URL: https://proceedings.mlr.press/v232/hemati23b.html PDF: https://proceedings.mlr.press/v232/hemati23b/hemati23b.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-hemati23b.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Hamed family: Hemati - given: Andrea family: Cossu - given: Antonio family: Carta - given: Julio family: Hurtado - given: Lorenzo family: Pellegrini - given: Davide family: Bacciu - given: Vincenzo family: Lomonaco - given: Damian family: Borth editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 437-455 id: hemati23b issued: date-parts: - 2023 - 11 - 20 firstpage: 437 lastpage: 455 published: 2023-11-20 00:00:00 +0000 - title: 'Reducing Communication Overhead in Federated Learning for Pre-trained Language Models Using Parameter-Efficient Finetuning' abstract: 'Pre-trained language models are shown to be effective in solving real-world natural language problems. Due to privacy reasons, data may not always be available for pre-training or finetuning of the model. Federated learning (FL) is a privacy-preserving technique for model training, but it suffers from communication overhead when the model size is large. We show that parameter-efficient finetuning (PEFT) reduces communication costs while achieving good model performance in both supervised and semi-supervised federated learning. Also, often in real life, data for the target downstream task is not available, but it is relatively easy to obtain the data for other related tasks. To this end, our results on the task-level transferability of PEFT methods in federated learning show that the model achieves good zero-shot performance on target data when source data is from a similar task. Parameter-efficient finetuning can aid federated learning in building efficient, privacy-preserving Natural Language Processing (NLP) applications.' volume: 232 URL: https://proceedings.mlr.press/v232/malaviya23a.html PDF: https://proceedings.mlr.press/v232/malaviya23a/malaviya23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-malaviya23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Shubham family: Malaviya - given: Manish family: Shukla - given: Sachin family: Lodha editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 456-469 id: malaviya23a issued: date-parts: - 2023 - 11 - 20 firstpage: 456 lastpage: 469 published: 2023-11-20 00:00:00 +0000 - title: 'Time and temporal abstraction in continual learning: tradeoffs, analogies and regret in an active measuring setting' abstract: 'This conceptual paper provides theoretical results linking notions in semi-supervised learning (SSL) and hierarchical reinforcement learning (HRL) in the context of lifelong learning. Specifically, our construction sets up a direct analogy between intermediate representations in SSL and temporal abstraction in RL, highlighting the important role of factorization in both types of hierarchy and the relevance of partial labeling, resp. partial observation. The construction centres around a simple class of Partially Observed Markov Decision Processes (POMDPs) where we show tools and results from SSL imply lower bounds on regret holding for any RL algorithm without access to temporal abstraction. While our lower bound is for a restricted class of RL problems, it applies to arbitrary RL algorithms in this setting. The setting moreover features so-called “active measuring”, an aspect of widespread relevance in industrial control, but - possibly due to its lifelong learning flavour - not yet well-studied in RL. Our formalization makes it possible to think about tradeoffs that apply for such control problems.' volume: 232 URL: https://proceedings.mlr.press/v232/letourneau23a.html PDF: https://proceedings.mlr.press/v232/letourneau23a/letourneau23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-letourneau23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Vincent family: Létourneau - given: Colin family: Bellinger - given: Isaac family: Tamblyn - given: Maia family: Fraser editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 470-480 id: letourneau23a issued: date-parts: - 2023 - 11 - 20 firstpage: 470 lastpage: 480 published: 2023-11-20 00:00:00 +0000 - title: 'Self-trained Centroid Classifiers for Semi-supervised Cross-domain Few-shot Learning' abstract: 'State-of-the-art cross-domain few-shot learning methods for image classification apply knowledge transfer by fine-tuning deep feature extractors obtained from source domains on the small labelled dataset available for the target domain, generally in conjunction with a simple centroid-based classification head. Semi-supervised learning during the meta-test phase is an obvious approach to incorporating unlabelled data into cross-domain few-shot learning, but semi-supervised methods designed for larger sets of labelled data than those available in few-shot learning appear to easily go astray when applied in this setting. We propose an efficient semi-supervised learning method that applies self-training to the classification head only and show that it can yield very consistent improvements in average performance in the Meta-Dataset benchmark for cross-domain few-shot learning when applied with contemporary methods utilising centroid-based classification.' volume: 232 URL: https://proceedings.mlr.press/v232/wang23a.html PDF: https://proceedings.mlr.press/v232/wang23a/wang23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-wang23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Hongyu family: Wang - given: Eibe family: Frank - given: Bernhard family: Pfahringer - given: Geoffrey family: Holmes editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 481-492 id: wang23a issued: date-parts: - 2023 - 11 - 20 firstpage: 481 lastpage: 492 published: 2023-11-20 00:00:00 +0000 - title: 'Evaluating Continual Learning on a Home Robot' abstract: 'Robots in home environments need to be able to learn new skills continuously as data becomes available, becoming ever more capable over time while using as little real-world data as possible. However, traditional robot learning approaches typically assume large amounts of iid data, which is inconsistent with this goal. In contrast, continuous learning methods like CLEAR and SANE allow autonomous agents to learn off of a stream of non-iid samples; they, however, have not previously been demonstrated on real robotics platforms. In this work, we show how continuous learning methods can be adapted for use on a real, low-cost home robot, and in particular look at the case where we have extremely small numbers of examples, in a task-id-free setting. Specifically, we propose SANER, a method for continuously learning a library of skills, and \model{} (Attention-Based PointNet) as the backbone to support it. We learn four sequential kitchen tasks on a low-cost home robot, using only a handful of demonstrations per task.' volume: 232 URL: https://proceedings.mlr.press/v232/powers23a.html PDF: https://proceedings.mlr.press/v232/powers23a/powers23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-powers23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Sam family: Powers - given: Abhinav family: Gupta - given: Chris family: Paxton editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 493-512 id: powers23a issued: date-parts: - 2023 - 11 - 20 firstpage: 493 lastpage: 512 published: 2023-11-20 00:00:00 +0000 - title: 'Fixed Design Analysis of Regularization-Based Continual Learning' abstract: 'We consider a continual learning (CL) problem with two linear regression tasks in the fixed design setting, where the feature vectors are assumed fixed and the labels are assumed to be random variables. We consider an $\ell_2$-regularized CL algorithm, which computes an Ordinary Least Squares parameter to fit the first dataset, then computes another parameter that fits the second dataset under an $\ell_2$-regularization penalizing its deviation from the first parameter, and outputs the second parameter. For this algorithm, we provide tight upper and lower bounds on the average risk over the two tasks. Our risk bounds reveal a provable trade-off between forgetting and intransigence of the $\ell_2$-regularized CL algorithm: with a large regularization parameter, the algorithm output forgets less information about the first task but is intransigent to extract new information from the second task; and vice versa. Our results suggest that catastrophic forgetting could happen for CL with dissimilar tasks (under a precise similarity measurement), and that a well-tuned $\ell_2$-regularization can partially mitigate this issue by introducing intransigence.' volume: 232 URL: https://proceedings.mlr.press/v232/li23b.html PDF: https://proceedings.mlr.press/v232/li23b/li23b.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-li23b.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Haoran family: Li - given: Jingfeng family: Wu - given: Vladimir family: Braverman editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 513-533 id: li23b issued: date-parts: - 2023 - 11 - 20 firstpage: 513 lastpage: 533 published: 2023-11-20 00:00:00 +0000 - title: 'Continually learning representations at scale' abstract: 'Many widely used continual learning benchmarks follow a protocol that starts from an untrained, randomly initialized model that needs to sequentially learn a number of incoming tasks. To maximize interpretability of the results and to keep experiment length under control, often these tasks are formed from well-known medium to large size datasets such as CIFAR or ImageNet. Recently, however, large-scale pretrained representations, also referred to as foundation models, have achieved significant success across a wide range of traditional vision and language problems. Furthermore, the availability of these pretrained models and their use as starting point for training can be seen as a paradigm shift from the classical end-to-end learning. This raises the question: How does this paradigm shift influence continual learning research? We attempt an answer, by firstly showing that many existing benchmarks are ill-equipped in this setting. The use of foundation model leads to state-of-art results on several existing and commonly used image classification continual learning benchmarks, from split CIFAR-100 to split ImageNet. Additionally, there is at best a small gap between keeping the representations frozen versus tuning them. While this is indicative of the overlap between pretraining distribution and the benchmark distribution, it also shows that these benchmarks can not be used to explore how to continually learn the underlying representations. Secondly, we examine what differentiates continually learning from scratch versus when relying on pretrained models, where the representation is learned under a different objective. We highlight that this brings about new challenges and research questions that cannot be studied in the sanitised scenario of learning from scratch explored so far. ' volume: 232 URL: https://proceedings.mlr.press/v232/galashov23a.html PDF: https://proceedings.mlr.press/v232/galashov23a/galashov23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-galashov23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Alexandre family: Galashov - given: Jovana family: Mitrovic - given: Dhruva family: Tirumala - given: Yee Whye family: Teh - given: Timothy family: Nguyen - given: Arslan family: Chaudhry - given: Razvan family: Pascanu editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 534-547 id: galashov23a issued: date-parts: - 2023 - 11 - 20 firstpage: 534 lastpage: 547 published: 2023-11-20 00:00:00 +0000 - title: 'Minimal Value-Equivalent Partial Models for Scalable and Robust Planning in Lifelong Reinforcement Learning' abstract: 'Learning models of the environment from pure interaction is often considered an essential component of building lifelong reinforcement learning agents. However, the common practice in model-based reinforcement learning is to learn models that model every aspect of the agent’s environment, regardless of whether they are important in coming up with optimal decisions or not. In this paper, we argue that such models are not particularly well-suited for performing scalable and robust planning in lifelong reinforcement learning scenarios and we propose new kinds of models that only model the relevant aspects of the environment, which we call \emph{minimal value-equivalent partial models}. After providing a formal definition for these models, we provide theoretical results demonstrating the scalability advantages of performing planning with such models and then perform experiments to empirically illustrate our theoretical results. Then, we provide some useful heuristics on how to learn these kinds of models with deep learning architectures and empirically demonstrate that models learned in such a way can allow for performing planning that is robust to distribution shifts and compounding model errors. Overall, both our theoretical and empirical results suggest that minimal value-equivalent partial models can provide significant benefits to performing scalable and robust planning in lifelong reinforcement learning scenarios. ' volume: 232 URL: https://proceedings.mlr.press/v232/alver23a.html PDF: https://proceedings.mlr.press/v232/alver23a/alver23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-alver23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Safa family: Alver - given: Doina family: Precup editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 548-567 id: alver23a issued: date-parts: - 2023 - 11 - 20 firstpage: 548 lastpage: 567 published: 2023-11-20 00:00:00 +0000 - title: 'Hierarchical Representation Learning for Markov Decision Processes' abstract: 'In this paper, we present a novel method for learning reward-agnostic hierarchical representations of Markov Decision Processes. Our method works by partitioning the state space into subsets, and defines subtasks for performing transitions between the partitions. At the high level, we use model-based planning to decide which subtask to pursue next from a given partition. We formulate the problem of partitioning the state space as an optimization problem that can be solved using gradient descent given a set of sampled trajectories, making our method suitable for high-dimensional problems with large state spaces. We empirically validate the method, by showing that it can successfully learn useful hierarchical representations in domains with high-dimensional states. Once learned, the hierarchical representation can be used to solve different tasks in the given domain, thus generalizing knowledge across tasks.' volume: 232 URL: https://proceedings.mlr.press/v232/steccanella23a.html PDF: https://proceedings.mlr.press/v232/steccanella23a/steccanella23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-steccanella23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Lorenzo family: Steccanella - given: Simone family: Totaro - given: Anders family: Jonsson editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 568-585 id: steccanella23a issued: date-parts: - 2023 - 11 - 20 firstpage: 568 lastpage: 585 published: 2023-11-20 00:00:00 +0000 - title: 'MultiMix TFT: A Multi-task Mixed-Frequency Framework with Temporal Fusion Transformers' abstract: 'Multi-task learning (MTL) has been increasingly recognized as an effective paradigm in time-series analysis for forecasting multiple related tasks concurrently. Prior MTL frameworks for time-series forecasting have typically been devised for tasks that share the same regular time frequencies. However, numerous real-world scenarios entail tasks measured at mixed, and often irregular, time frequencies. We propose a multi-task mixed-frequency (MultiMix) learning framework for time-series forecasting that addresses the challenges of mixed-frequency scenarios where tasks are measured at different and/or irregular time intervals. Our proposed framework leverages the relationships between mixed-frequency tasks to improve accuracy and robustness of time-series forecasting across tasks. The MultiMix framework is implemented using the state-of-the-art Temporal Fusion Transformer (TFT) and is evaluated in smart irrigation, where predicting mid-day stem water potential and soil water potential pose critical challenges. The MultiMix TFT enables joint forecasting of stem water potential, measured sparsely on irregular and infrequent time intervals, and soil water potential, measured on a daily time interval. The results show substantial improvements in stem water potential prediction over state-of-the-art baselines while achieving comparable performance for soil water potential. These results confirm the effectiveness of the proposed framework for addressing the mixed-frequency time-series forecasting problem in real-world settings. Code will be made available upon publication.' volume: 232 URL: https://proceedings.mlr.press/v232/deforce23a.html PDF: https://proceedings.mlr.press/v232/deforce23a/deforce23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-deforce23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Boje family: Deforce - given: Bart family: Baesens - given: Jan family: Diels - given: Estefanía Serral family: Asensio editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 586-600 id: deforce23a issued: date-parts: - 2023 - 11 - 20 firstpage: 586 lastpage: 600 published: 2023-11-20 00:00:00 +0000 - title: 'What Happens During Finetuning of Vision Transformers: An Invariance Based Investigation' abstract: 'The pretrain-finetune paradigm usually improves downstream performance over training a model from scratch on the same task, becoming commonplace across many areas of machine learning. While pretraining is empirically observed to be beneficial for a range of tasks, there is not a clear understanding yet of the reasons for this effect. In this work, we examine the relationship between pretrained vision transformers and the corresponding finetuned versions on several benchmark datasets and tasks. We present new metrics that specifically investigate the degree to which invariances learned by a pretrained model are retained or forgotten during finetuning. Using these metrics, we present a suite of empirical findings, including that pretraining induces transferable invariances in shallow layers and that invariances from deeper pretrained layers are compressed towards shallower layers during finetuning. Together, these findings contribute to understanding some of the reasons for the successes of pretrained models and the changes that a pretrained model undergoes when finetuned on a downstream task. ' volume: 232 URL: https://proceedings.mlr.press/v232/merlin23a.html PDF: https://proceedings.mlr.press/v232/merlin23a/merlin23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-merlin23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Gabriele family: Merlin - given: Vedant family: Nanda - given: Ruchit family: Rawal - given: Mariya family: Toneva editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 601-619 id: merlin23a issued: date-parts: - 2023 - 11 - 20 firstpage: 601 lastpage: 619 published: 2023-11-20 00:00:00 +0000 - title: 'Loss of Plasticity in Continual Deep Reinforcement Learning' abstract: 'In this paper, we characterize the behavior of canonical value-based deep reinforcement learning (RL) approaches under varying degrees of non-stationarity. In particular, we demonstrate that deep RL agents lose their ability to learn good policies when they cycle through a sequence of Atari 2600 games. This phenomenon is alluded to in prior work under various guises—e.g., loss of plasticity, implicit under-parameterization, primacy bias, and capacity loss. We investigate this phenomenon closely at scale and analyze how the weights, gradients, and activations change over time in several experiments with varying experimental conditions (e.g., similarity between games, number of games, number of frames per game). Our analysis shows that the activation footprint of the network becomes sparser, contributing to the diminishing gradients. We investigate a remarkably simple mitigation strategy—Concatenated ReLUs (CReLUs) activation function—and demonstrate its effectiveness in facilitating continual learning in a changing environment.' volume: 232 URL: https://proceedings.mlr.press/v232/abbas23a.html PDF: https://proceedings.mlr.press/v232/abbas23a/abbas23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-abbas23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Zaheer family: Abbas - given: Rosie family: Zhao - given: Joseph family: Modayil - given: Adam family: White - given: Marlos C. family: Machado editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 620-636 id: abbas23a issued: date-parts: - 2023 - 11 - 20 firstpage: 620 lastpage: 636 published: 2023-11-20 00:00:00 +0000 - title: 'Sample-Efficient Learning of Novel Visual Concepts' abstract: 'Despite the advances made in visual object recognition, state-of-the-art deep learning models struggle to effectively recognize novel objects in a few-shot setting where only a limited number of examples are provided. Unlike humans who excel at such tasks, these models often fail to leverage known relationships between entities in order to draw conclusions about such objects. In this work, we show that incorporating a symbolic knowledge graph into a state-of-the-art recognition model enables a new approach for effective few-shot classification. In our proposed neuro-symbolic architecture and training methodology, the knowledge graph is augmented with additional relationships extracted from a small set of examples, improving its ability to recognize novel objects by considering the presence of interconnected entities. Unlike existing few-shot classifiers, we show that this enables our model to incorporate not only objects but also abstract concepts and affordances. The existence of the knowledge graph also makes this approach amenable to interpretability through analysis of the relationships contained within it. We empirically show that our approach outperforms current state-of-the-art few-shot multi-label classification methods on the COCO dataset and evaluate the addition of abstract concepts and affordances on the Visual Genome dataset. ' volume: 232 URL: https://proceedings.mlr.press/v232/bhagat23a.html PDF: https://proceedings.mlr.press/v232/bhagat23a/bhagat23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-bhagat23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Sarthak family: Bhagat - given: Simon family: Stepputtis - given: Joseph family: Campbell - given: Katia family: Sycara editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 637-657 id: bhagat23a issued: date-parts: - 2023 - 11 - 20 firstpage: 637 lastpage: 657 published: 2023-11-20 00:00:00 +0000 - title: 'VIBR: Learning View-Invariant Value Functions for Robust Visual Control' abstract: 'End-to-end reinforcement learning on images showed significant progress in the recent years. Data-based approach leverage data augmentation and domain randomization while representation learning methods use auxiliary losses to learn task-relevant features. Yet, reinforcement still struggles in visually diverse environments full of distractions and spurious noise. In this work, we tackle the problem of robust visual control at its core and present VIBR (View-Invariant Bellman Residuals), a method that combines multi-view training and invariant prediction to reduce out-of-distribution (OOD) generalization gap for RL based visuomotor control. Our model-free approach improve baselines performances without the need of additional representation learning objectives and with limited additional computational cost. We show that VIBR outperforms existing methods on complex visuo-motor control environment with high visual perturbation. Our approach achieves state-of the-art results on the Distracting Control Suite benchmark, a challenging benchmark still not solved by current methods, where we evaluate the robustness to a number of visual perturbators, as well as OOD generalization and extrapolation capabilities.' volume: 232 URL: https://proceedings.mlr.press/v232/dupuis23a.html PDF: https://proceedings.mlr.press/v232/dupuis23a/dupuis23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-dupuis23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Tom family: Dupuis - given: Jaonary family: Rabarisoa - given: Quoc-Cuong family: Pham - given: David family: Filliat editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 658-682 id: dupuis23a issued: date-parts: - 2023 - 11 - 20 firstpage: 658 lastpage: 682 published: 2023-11-20 00:00:00 +0000 - title: 'Incremental Unsupervised Domain Adaptation on Evolving Graphs' abstract: 'Non-stationary data distributions in evolving graphs can create problems for deployed graph neural networks (GNN), such as fraud detection GNNs that can become ineffective when fraudsters alter their patterns. The aim of this study is to investigate how to incrementally adapt graph neural networks to incoming, unlabeled graph data after training and deployment. To achieve this, we propose a new approach called graph contrastive self-training (GCST) that combines contrastive learning and self-training to alleviate performance drop. To evaluate the effectiveness of our approach, we conduct a comprehensive empirical evaluation on four diverse graph datasets, comparing it to domain-invariant feature learning methods and plain self-training methods. Our contribution is three-fold: we formulate and study incremental unsupervised domain adaptation on evolving graphs, present an approach that integrates contrastive learning and self-training, and conduct a comprehensive empirical evaluation of our approach, which demonstrates its stability and superiority over other methods.' volume: 232 URL: https://proceedings.mlr.press/v232/chung23a.html PDF: https://proceedings.mlr.press/v232/chung23a/chung23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-chung23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Hsing-Huan family: Chung - given: Joydeep family: Ghosh editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 683-702 id: chung23a issued: date-parts: - 2023 - 11 - 20 firstpage: 683 lastpage: 702 published: 2023-11-20 00:00:00 +0000 - title: 'Auxiliary task discovery through generate-and-test' abstract: 'In this paper, we explore an approach to auxiliary task discovery in reinforcement learning based on ideas from representation learning. Auxiliary tasks tend to improve data efficiency by forcing the agent to learn auxiliary prediction and control objectives in addition to the main task of maximizing reward, and thus producing better representations. Typically these tasks are designed by people. Meta-learning offers a promising avenue for automatic task discovery; however, these methods are computationally expensive and challenging to tune in practice. In this paper, we explore a complementary approach to the auxiliary task discovery: continually generating new auxiliary tasks and preserving only those with high utility. We also introduce a new measure of auxiliary tasks’ usefulness based on how useful the features induced by them are for the main task. Our discovery algorithm significantly outperforms random tasks and learning without auxiliary tasks across a suite of environments. ' volume: 232 URL: https://proceedings.mlr.press/v232/rafiee23a.html PDF: https://proceedings.mlr.press/v232/rafiee23a/rafiee23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-rafiee23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Banafsheh family: Rafiee - given: Sina family: Ghiassian - given: Jun family: Jin - given: Richard family: Sutton - given: Jun family: Luo - given: Adam family: White editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 703-714 id: rafiee23a issued: date-parts: - 2023 - 11 - 20 firstpage: 703 lastpage: 714 published: 2023-11-20 00:00:00 +0000 - title: 'Restarted Bayesian Online Change-point Detection for Non-Stationary Markov Decision Processes' abstract: 'We consider the problem of learning in a non-stationary reinforcement learning (RL) environment, where the setting can be fully described by a piecewise stationary discrete-time Markov decision process (MDP). We introduce a variant of the Restarted Bayesian Online Change-Point Detection algorithm (R-BOCPD) that operates on input streams originating from the more general multinomial distribution and provides near-optimal theoretical guarantees in terms of false-alarm rate and detection delay. Based on this, we propose an improved version of the UCRL2 algorithm for MDPs with state transition kernel sampled from a multinomial distribution, which we call R-BOCPD-UCRL2. We perform a finite-time performance analysis and show that R-BOCPD-UCRL2 enjoys a favorable regret bound of $\mathcal{O}\left(D O \sqrt{A T K_T \log\left (\frac{T}{\delta} \right)} + \frac{K_T \log \frac{K_T}{\delta}}{\min\limits_\ell \:{KL}(\boldsymbol{\theta}^{(\ell+1)},\boldsymbol{\theta}^{(\ell)})} \right)$, where $D$ is the largest MDP diameter from the set of MDPs defining the piecewise stationary MDP setting, $O$ is the finite number of states (constant over all changes), $A$ is the finite number of actions (constant over all changes), $K_T$ is the number of change points, and $\boldsymbol{\theta}^{(\ell)}$ is the transition kernel during the interval $[c_\ell, c_{\ell+1})$, which we assume to be multinomially distributed over the set of states $\mathsf{O}$. Interestingly, the performance bound does not directly scale with the variation in MDP state transition distributions and rewards, ie. can also model abrupt changes. In practice, R-BOCPD-UCRL2 outperforms the state-of-the-art in a variety of scenarios in synthetic environments. We provide a detailed experimental setup along with a code repository (upon publication) that can be used to easily reproduce our experiments.' volume: 232 URL: https://proceedings.mlr.press/v232/alami23a.html PDF: https://proceedings.mlr.press/v232/alami23a/alami23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-alami23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Reda family: Alami - given: Mohammed family: Mahfoud - given: Eric family: Moulines editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 715-744 id: alami23a issued: date-parts: - 2023 - 11 - 20 firstpage: 715 lastpage: 744 published: 2023-11-20 00:00:00 +0000 - title: 'Value-aware Importance Weighting for Off-policy Reinforcement Learning' abstract: 'Importance sampling is a central idea underlying off-policy prediction in reinforcement learning. It provides a strategy for re-weighting samples from a distribution to represent unbiased estimates of another distribution. However, importance sampling weights tend to be of high variance, often leading to stability issues in practice. In this work, we consider a broader class of importance weights to correct samples in off-policy learning. We propose the use of value-aware importance weights which take into account the sample space to provide lower variance, but still unbiased, estimates under a target distribution. We derive how such weights can be computed, and detail key properties of the resulting importance weights. We then extend several reinforcement learning prediction algorithms to the off-policy setting with these weights, and evaluate them empirically.' volume: 232 URL: https://proceedings.mlr.press/v232/de-asis23a.html PDF: https://proceedings.mlr.press/v232/de-asis23a/de-asis23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-de-asis23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Kristopher family: De Asis - given: Eric family: Graves - given: Richard S. family: Sutton editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 745-763 id: de-asis23a issued: date-parts: - 2023 - 11 - 20 firstpage: 745 lastpage: 763 published: 2023-11-20 00:00:00 +0000 - title: 'Re-Weighted Softmax Cross-Entropy to Control Forgetting in Federated Learning' abstract: 'In Federated Learning a global model is learned by aggregating model updates computed at a set of independent client nodes. To reduce communication costs, multiple gradient steps are performed at each node prior to aggregation. A key challenge in this setting is data heterogeneity across clients resulting in differing local objectives. This can lead clients to overly minimize their own local objective consequently diverging from the global solution. We demonstrate that individual client models experience a catastrophic forgetting with respect to data from other clients and propose an efficient approach that modifies the cross-entropy objective on a per-client basis by re-weighting the softmax logits prior to computing the loss. This approach shields classes outside a client’s label set from abrupt representation change and we empirically demonstrate it can alleviate client forgetting and provide consistent improvements to standard federated learning algorithms. Our method is particularly beneficial under the most challenging federated learning settings where data heterogeneity is high and client participation in each round is low.' volume: 232 URL: https://proceedings.mlr.press/v232/legate23a.html PDF: https://proceedings.mlr.press/v232/legate23a/legate23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-legate23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Gwen family: Legate - given: Lucas family: Caccia - given: Eugene family: Belilovsky editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 764-780 id: legate23a issued: date-parts: - 2023 - 11 - 20 firstpage: 764 lastpage: 780 published: 2023-11-20 00:00:00 +0000 - title: 'Measuring and Mitigating Interference in Reinforcement Learning' abstract: 'Catastrophic interference is common in many network-based learning systems, and many proposals exist for mitigating it. Before overcoming interference we must understand it better. In this work, we provide a definition and novel measure of interference for value-based reinforcement learning methods such as Fitted Q-Iteration and DQN. We systematically evaluate our measure of interference, showing that it correlates with instability in control performance, across a variety of network architectures. Our new interference measure allows us to ask novel scientific questions about commonly used deep learning architectures and study learning algorithms which mitigate interference. Lastly, we outline a class of algorithms which we call online-aware that are designed to mitigate interference, and show they do reduce interference according to our measure and that they improve stability and performance in several classic control environments.' volume: 232 URL: https://proceedings.mlr.press/v232/liu23a.html PDF: https://proceedings.mlr.press/v232/liu23a/liu23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-liu23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Vincent family: Liu - given: Han family: Wang - given: Ruo Yu family: Tao - given: Khurram family: Javed - given: Adam family: White - given: Martha family: White editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 781-795 id: liu23a issued: date-parts: - 2023 - 11 - 20 firstpage: 781 lastpage: 795 published: 2023-11-20 00:00:00 +0000 - title: 'Adaptive Meta-Learning via data-dependent PAC-Bayes bounds' abstract: 'Meta-learning aims to extract common knowledge from similar training tasks in order to facilitate efficient and effective learning on future tasks. Several recent works have extended PAC-Bayes generalization error bounds to the meta-learning setting. By doing so, prior knowledge can be incorporated in the form of a distribution over hypotheses that is expected to lead to low error on new tasks that are similar to those that have been previously observed. In this work, we develop novel bounds for the generalization error on test tasks based on recent data-dependent bounds and provide a novel algorithm for adapting prior knowledge to downstream tasks in a potentially more effective manner. We demonstrate the effectiveness of our algorithm numerically for few-shot image classification tasks with deep neural networks and show a significant reduction in generalization error without any additional adaptation data.' volume: 232 URL: https://proceedings.mlr.press/v232/friedman23a.html PDF: https://proceedings.mlr.press/v232/friedman23a/friedman23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-friedman23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Lior family: Friedman - given: Ron family: Meir editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 796-810 id: friedman23a issued: date-parts: - 2023 - 11 - 20 firstpage: 796 lastpage: 810 published: 2023-11-20 00:00:00 +0000 - title: 'Active Class Selection for Few-Shot Class-Incremental Learning' abstract: 'For real-world applications, robots will need to continually learn in their environments through limited interactions with their users. Toward this, previous works in few-shot class incremental learning (FSCIL) and active class selection (ACS) have achieved promising results but were tested in constrained setups. Therefore, in this paper, we combine ideas from FSCIL and ACS to develop a novel framework that can allow an autonomous agent to continually learn new objects by asking its users to label only a few of the most informative objects in the environment. To this end, we build on a state-of-the-art (SOTA) FSCIL model and extend it with techniques from ACS literature. We term this model Few-shot Incremental Active class SeleCtiOn (FIASco). We further integrate a potential field-based navigation technique with our model to develop a complete framework that can allow an agent to process and reason on its sensory data through the FIASco model, navigate towards the most informative object in the environment, gather data about the object through its sensors and incrementally update the FIASco model. Experimental results on a simulated agent and a real robot show the significance of our approach for long-term real-world robotics applications.' volume: 232 URL: https://proceedings.mlr.press/v232/mcclurg23a.html PDF: https://proceedings.mlr.press/v232/mcclurg23a/mcclurg23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-mcclurg23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Christopher family: McClurg - given: Ali family: Ayub - given: Harsh family: Tyagi - given: Sarah M. family: Rajtmajer - given: Alan R. family: Wagner editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 811-827 id: mcclurg23a issued: date-parts: - 2023 - 11 - 20 firstpage: 811 lastpage: 827 published: 2023-11-20 00:00:00 +0000 - title: 'Improving Online Continual Learning Performance and Stability with Temporal Ensembles' abstract: 'Neural networks are very effective when trained on large datasets for a large number of iterations. However, when they are trained on non-stationary streams of data and in an online fashion, their performance is reduced (1) by the online setup, which limits the availability of data, (2) due to catastrophic forgetting because of the non-stationary nature of the data. Furthermore, several recent works (Caccia et al. 2022, Lange et al. 2023) showed that replay methods used in continual learning suffer from the \textit{stability gap}, encountered when evaluating the model continually (rather than only on task boundaries). In this article, we study the effect of model ensembling as a way to improve performance and stability in online continual learning. We notice that naively ensembling models coming from a variety of training tasks increases the performance in online continual learning considerably. Starting from this observation, and drawing inspirations from semi-supervised learning ensembling methods, we use a lightweight temporal ensemble that computes the exponential moving average of the weights (EMA) at test time, and show that it can drastically increase the performance and stability when used in combination with several methods from the literature.' volume: 232 URL: https://proceedings.mlr.press/v232/soutif-cormerais23a.html PDF: https://proceedings.mlr.press/v232/soutif-cormerais23a/soutif-cormerais23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-soutif-cormerais23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Albin family: Soutif–Cormerais - given: Antonio family: Carta - given: Joost prefix: van de family: Weijer editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 828-845 id: soutif-cormerais23a issued: date-parts: - 2023 - 11 - 20 firstpage: 828 lastpage: 845 published: 2023-11-20 00:00:00 +0000 - title: 'Model-Based Meta Automatic Curriculum Learning' abstract: 'Curriculum learning (CL) has been widely explored to facilitate the learning of hard-exploration tasks in reinforcement learning (RL) by training a sequence of easier tasks, often called a curriculum. While most curricula are built either manually or automatically based on heuristics, e.g. choosing a training task which is barely beyond the current abilities of the learner, the fact that similar tasks might benefit from similar curricula motivates us to explore meta-learning as a technique for curriculum generation or teaching for a distribution of similar tasks. This paper formulates the meta CL problem that requires a meta-teacher to generate the curriculum which will assist the student to train toward any given target task from a task distribution based on the similarity of these tasks to one another. We propose a model-based meta automatic curriculum learning algorithm (MM-ACL) that learns to predict the performance improvement on one task when the student is trained on another, given the current status of the student. This predictor can then be used to generate the curricula for different target tasks. Our empirical results demonstrate that MM-ACL outperforms the state-of-the-art CL algorithms in a grid-world domain and a more complex visual-based navigation domain in terms of sample efficiency.' volume: 232 URL: https://proceedings.mlr.press/v232/xu23a.html PDF: https://proceedings.mlr.press/v232/xu23a/xu23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-xu23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Zifan family: Xu - given: Yulin family: Zhang - given: Shahaf S. family: Shperberg - given: Reuth family: Mirsky - given: Yuqian family: Jiang - given: Bo family: Liu - given: Peter family: Stone editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 846-860 id: xu23a issued: date-parts: - 2023 - 11 - 20 firstpage: 846 lastpage: 860 published: 2023-11-20 00:00:00 +0000 - title: 'Towards Few-shot Coordination: Revisiting Ad-hoc Teamplay Challenge In the Game of Hanabi' abstract: 'Cooperative Multi-agent Reinforcement Learning (MARL) algorithms with Zero-Shot Coordination (ZSC) have gained significant attention in recent years. ZSC refers to the ability of agents to coordinate with independently trained agents. While ZSC is crucial for cooperative MARL agents, it might not be possible for complex tasks and changing environments. Agents also need to adapt and improve their performance with minimal interaction with other agents. In this work, we show empirically that state-of-the-art ZSC algorithms have poor performance when paired with agents trained with different methods, and they require millions of samples to adapt to these new partners. To investigate this issue, we formally defined a framework based on a popular cooperative multi-agent game called Hanabi to evaluate the adaptability of MARL methods. In particular, we created a diverse set of pre-trained agents and defined a new metric called adaptation regret that measures the agent’s ability to efficiently adapt and improve its coordination performance when paired with some held-out pool of partners on top of its ZSC performance. After evaluating several SOTA algorithms using our framework, our experiments reveal that naive Independent Q-Learning (IQL) agents in most cases adapt as quickly as the SOTA ZSC algorithm Off-Belief Learning (OBL). This finding raises an interesting research question: How to design MARL algorithms with high ZSC performance and capability of fast adaptation to unseen partners. As a first step, we studied the role of different hyper-parameters and design choices on the adaptability of current MARL algorithms. Our experiments show that two categories of hyper-parameters controlling the data diversity and optimization process have a significant impact on the adaptability of Hanabi agents. We hope this initial analysis will inspire more work on designing both general and adaptive MARL algorithms.' volume: 232 URL: https://proceedings.mlr.press/v232/nekoei23b.html PDF: https://proceedings.mlr.press/v232/nekoei23b/nekoei23b.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-nekoei23b.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Hadi family: Nekoei - given: Xutong family: Zhao - given: Janarthanan family: Rajendran - given: Miao family: Liu - given: Sarath family: Chandar editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 861-877 id: nekoei23b issued: date-parts: - 2023 - 11 - 20 firstpage: 861 lastpage: 877 published: 2023-11-20 00:00:00 +0000 - title: 'Comparing the Efficacy of Fine-Tuning and Meta-Learning for Few-Shot Policy Imitation' abstract: 'In this paper we explore few-shot imitation learning for control problems, which involves learning to imitate a target policy by accessing a limited set of offline rollouts. This setting has been relatively under-explored despite its relevance to robotics and control applications. State-of-the-art methods developed to tackle few-shot imitation rely on meta-learning, which is expensive to train as it requires access to a distribution over tasks (rollouts from many target policies and variations of the base environment). Given this limitation we investigate an alternative approach, fine-tuning, a family of methods that pretrain on a single dataset and then fine-tune on unseen domain-specific data. Recent work has shown that fine-tuners outperform meta-learners in few-shot image classification tasks, especially when the data is out-of-domain. Here we evaluate to what extent this is true for control problems, proposing a simple yet effective baseline which relies on two stages: (i) training a base policy online via reinforcement learning (e.g. Soft Actor-Critic) on a single base environment, (ii) fine-tuning the base policy via behavioral cloning on a few offline rollouts of the target policy. Despite its simplicity this baseline is competitive with meta-learning methods on a variety of conditions and is able to imitate target policies trained on unseen variations of the original environment. Importantly, the proposed approach is practical and easy to implement, as it does not need any complex meta-training protocol. As a further contribution, we release an open source dataset called iMuJoCo (iMitation MuJoCo) consisting of 154 variants of popular OpenAI-Gym MuJoCo environments with associated pretrained target policies and rollouts, which can be used by the community to study few-shot imitation learning and offline reinforcement learning.' volume: 232 URL: https://proceedings.mlr.press/v232/patacchiola23a.html PDF: https://proceedings.mlr.press/v232/patacchiola23a/patacchiola23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-patacchiola23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Massimiliano family: Patacchiola - given: Mingfei family: Sun - given: Katja family: Hofmann - given: Richard E. family: Turner editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 878-908 id: patacchiola23a issued: date-parts: - 2023 - 11 - 20 firstpage: 878 lastpage: 908 published: 2023-11-20 00:00:00 +0000 - title: 'Substituting Data Annotation with Balanced Neighbourhoods and Collective Loss in Multi-label Text Classification' abstract: 'Multi-label text classification (MLTC) is the task of assigning multiple labels to a given text, and has a wide range of application domains. Most existing approaches require an enormous amount of annotated data to learn a classifier and/or a set of well-defined constraints on the label space structure, such as hierarchical relations which may be complicated to provide as the number of labels increases. In this paper, we study the MLTC problem in annotation-free and scarce-annotation settings in which the magnitude of available supervision signals is linear to the number of labels. Our method follows three steps, (1) mapping input text into a set of preliminary label likelihoods by natural language inference using a pre-trained language model, (2) calculating a signed label dependency graph by label descriptions, and (3) updating the preliminary label likelihoods with message passing along the label dependency graph, driven with a collective loss function that injects the information of expected label frequency and average multi-label cardinality of predictions. The experiments show that the proposed framework achieves effective performance under low supervision settings with almost imperceptible computational and memory overheads added to the usage of pre-trained language model outperforming its initial performance by 70% in terms of example-based F1 score. ' volume: 232 URL: https://proceedings.mlr.press/v232/ozmen23a.html PDF: https://proceedings.mlr.press/v232/ozmen23a/ozmen23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-ozmen23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Muberra family: Ozmen - given: Joseph family: Cotnareanu - given: Mark family: Coates editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 909-922 id: ozmen23a issued: date-parts: - 2023 - 11 - 20 firstpage: 909 lastpage: 922 published: 2023-11-20 00:00:00 +0000 - title: 'I2I: Initializing Adapters with Improvised Knowledge' abstract: 'Adapters present a promising solution to the catastrophic forgetting problem in continual learning. However, training independent Adapter modules for every new task misses an opportunity for cross-task knowledge transfer. We propose Improvise to Initialize (I2I), a continual learning algorithm that initializes Adapters for incoming tasks by distilling knowledge from previously-learned tasks’ Adapters. We evaluate I2I on CLiMB, a multimodal continual learning benchmark, by conducting experiments on sequences of visual question answering tasks. Adapters trained with I2I consistently achieve better task accuracy than independently-trained Adapters, demonstrating that our algorithm facilitates knowledge transfer between task Adapters. I2I also results in better cross-task knowledge transfer than the state-of-the-art AdapterFusion without incurring the associated parametric cost.' volume: 232 URL: https://proceedings.mlr.press/v232/srinivasan23a.html PDF: https://proceedings.mlr.press/v232/srinivasan23a/srinivasan23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-srinivasan23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Tejas family: Srinivasan - given: Furong family: Jia - given: Mohammad family: Rostami - given: Jesse family: Thomason editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 923-935 id: srinivasan23a issued: date-parts: - 2023 - 11 - 20 firstpage: 923 lastpage: 935 published: 2023-11-20 00:00:00 +0000 - title: 'Sharing Lifelong Reinforcement Learning Knowledge via Modulating Masks' abstract: 'Lifelong learning agents aim to learn multiple tasks sequentially over a lifetime. This involves the ability to exploit previous knowledge when learning new tasks and to avoid forgetting. Recently, modulating masks, a specific type of parameter isolation approach, have shown promise in both supervised and reinforcement learning. While lifelong learning algorithms have been investigated mainly within a single-agent approach, a question remains on how multiple agents can share lifelong learning knowledge with each other. We show that the parameter isolation mechanism used by modulating masks is particularly suitable for exchanging knowledge among agents in a distributed and decentralized system of lifelong learners. The key idea is that isolating specific task knowledge to specific masks allows agents to transfer only specific knowledge on-demand, resulting in a robust and effective collective of agents. We assume fully distributed and asynchronous scenarios with dynamic agent numbers and connectivity. An on-demand communication protocol ensures agents query their peers for specific masks to be transferred and integrated into their policies when facing each task. Experiments indicate that on-demand mask communication is an effective way to implement distributed and decentralized lifelong reinforcement learning, and provides a lifelong learning benefit with respect to distributed RL baselines such as DD-PPO, IMPALA, and PPO+EWC. The system is particularly robust to connection drops and demonstrates rapid learning due to knowledge exchange.' volume: 232 URL: https://proceedings.mlr.press/v232/nath23a.html PDF: https://proceedings.mlr.press/v232/nath23a/nath23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-nath23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Saptarshi family: Nath - given: Christos family: Peridis - given: Eseoghene family: Ben-Iwhiwhu - given: Xinran family: Liu - given: Shirin family: Dora - given: Cong family: Liu - given: Soheil family: Kolouri - given: Andrea family: Soltoggio editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 936-960 id: nath23a issued: date-parts: - 2023 - 11 - 20 firstpage: 936 lastpage: 960 published: 2023-11-20 00:00:00 +0000 - title: 'Continual Learning Beyond a Single Model' abstract: 'A growing body of research in continual learning focuses on the catastrophic forgetting problem. While many attempts have been made to alleviate this problem, the majority of the methods assume a \textit{single model} in the continual learning setup. In this work, we question this assumption and show that employing \textit{ensemble models} can be a simple yet effective method to improve continual performance. However, ensembles’ training and inference costs can increase significantly as the number of models grows. Motivated by this limitation, we study different ensemble models to understand their benefits and drawbacks in continual learning scenarios. Finally, to overcome the high compute cost of ensembles, we leverage recent advances in neural network subspace to propose a computationally cheap algorithm with similar runtime to a single model yet enjoying the performance benefits of ensembles.' volume: 232 URL: https://proceedings.mlr.press/v232/doan23a.html PDF: https://proceedings.mlr.press/v232/doan23a/doan23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-doan23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Thang family: Doan - given: Seyed Iman family: Mirzadeh - given: Mehrdad family: Farajtabar editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 961-991 id: doan23a issued: date-parts: - 2023 - 11 - 20 firstpage: 961 lastpage: 991 published: 2023-11-20 00:00:00 +0000 - title: 'Improving Performance in Continual Learning Tasks using Bio-Inspired Architectures' abstract: 'The ability to learn continuously from an incoming data stream without catastrophic forgetting is critical to designing intelligent systems. Many approaches to continual learning rely on stochastic gradient descent and its variants that employ global error updates, and hence need to adopt strategies such as memory buffers or replay to circumvent its stability, greed, and short-term memory limitations. To address this limitation, we have developed a biologically inspired lightweight neural network architecture that incorporates synaptic plasticity mechanisms and neuromodulation and hence learns through local error signals to enable online continual learning without stochastic gradient descent. Our approach leads to superior online continual learning performance on Split-MNIST, Split-CIFAR-10, and Split-CIFAR-100 datasets compared to other memory-constrained learning approaches and matches that of the state-of-the-art memory-intensive replay-based approaches. We further demonstrate the effectiveness of our approach by integrating key design concepts into other backpropagation-based continual learning algorithms, significantly improving their accuracy. Our results provide compelling evidence for the importance of incorporating biological principles into machine learning models and offer insights into how we can leverage them to design more efficient and robust systems for online continual learning.' volume: 232 URL: https://proceedings.mlr.press/v232/madireddy23a.html PDF: https://proceedings.mlr.press/v232/madireddy23a/madireddy23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-madireddy23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Sandeep family: Madireddy - given: Angel family: Yanguas-Gil - given: Prasanna family: Balaprakash editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 992-1008 id: madireddy23a issued: date-parts: - 2023 - 11 - 20 firstpage: 992 lastpage: 1008 published: 2023-11-20 00:00:00 +0000 - title: 'A Minimalist Approach for Domain Adaptation with Optimal Transport' abstract: 'We reveal an intriguing connection between adversarial attacks and cycle monotone maps, also known as optimal transport maps. Based on this finding, we developed a novel method named \textit{source fiction} for semi-supervised optimal transport-based domain adaptation. We conduct experiments on various datasets and show that our method can notably improve the performance of the optimal transport solvers in domain adaptation.' volume: 232 URL: https://proceedings.mlr.press/v232/asadulaev23a.html PDF: https://proceedings.mlr.press/v232/asadulaev23a/asadulaev23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-asadulaev23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Arip family: Asadulaev - given: Vitaly family: Shutov - given: Alexander family: Korotin - given: Alexander family: Panfilov - given: Vladislava family: Kontsevaya - given: Andrey family: Filchenkov editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 1009-1024 id: asadulaev23a issued: date-parts: - 2023 - 11 - 20 firstpage: 1009 lastpage: 1024 published: 2023-11-20 00:00:00 +0000 - title: 'Low-rank extended Kalman filtering for online learning of neural networks from streaming data' abstract: 'We propose an efficient online approximate Bayesian inference algorithm for estimating the parameters of a nonlinear function from a potentially non-stationary data stream. The method is based on the extended Kalman filter (EKF), but uses a novel low-rank plus diagonal decomposition of the posterior precision matrix, which gives a cost per step which is linear in the number of model parameters. In contrast to methods based on stochastic variational inference, our method is fully deterministic, and does not require step-size tuning. We show experimentally that this results in much faster (more sample efficient) learning, which results in more rapid adaptation to changing distributions, and faster accumulation of reward when used as part of a contextual bandit algorithm.' volume: 232 URL: https://proceedings.mlr.press/v232/chang23a.html PDF: https://proceedings.mlr.press/v232/chang23a/chang23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-chang23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Peter G. family: Chang - given: Gerardo family: Durán-Martín - given: Alex family: Shestopaloff - given: Matt family: Jones - given: Kevin Patrick family: Murphy editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 1025-1071 id: chang23a issued: date-parts: - 2023 - 11 - 20 firstpage: 1025 lastpage: 1071 published: 2023-11-20 00:00:00 +0000 - title: 'Introspective Action Advising for Interpretable Transfer Learning' abstract: 'Transfer learning can be applied in deep reinforcement learning to accelerate the training of a policy in a target task by transferring knowledge from a policy learned in a related source task. This is commonly achieved by copying pretrained weights from the source policy to the target policy prior to training, under the constraint that they use the same model architecture. However, not only does this require a robust representation learned over a wide distribution of states – often failing to transfer between specialist models trained over single tasks – but it is largely uninterpretable and provides little indication of what knowledge is transferred. In this work, we propose an alternative approach to transfer learning between tasks based on action advising, in which a teacher trained in a source task actively guides a student’s exploration in a target task. Through introspection, the teacher is capable of identifying when advice is beneficial to the student and should be given, and when it is not. Our approach allows knowledge transfer between policies agnostic of the underlying representations, and we empirically show that this leads to improved convergence rates in Gridworld and Atari environments while providing insight into what knowledge is transferred. ' volume: 232 URL: https://proceedings.mlr.press/v232/campbell23a.html PDF: https://proceedings.mlr.press/v232/campbell23a/campbell23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-campbell23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Joseph family: Campbell - given: Yue family: Guo - given: Fiona family: Xie - given: Simon family: Stepputtis - given: Katia family: Sycara editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 1072-1090 id: campbell23a issued: date-parts: - 2023 - 11 - 20 firstpage: 1072 lastpage: 1090 published: 2023-11-20 00:00:00 +0000 - title: 'SF-FSDA: Source-Free Few-Shot Domain Adaptive Object Detection with Efficient Labeled Data Factory' abstract: 'Domain adaptive object detection aims to leverage the knowledge learned from a labeled source domain to improve the performance on an unlabeled target domain. Prior works typically require the access to the source domain data for adaptation, and the availability of sufficient data on the target domain. However, these assumptions may not hold due to data privacy and rare data collection. In this paper, we propose and investigate a more practical and challenging domain adaptive object detection problem under both source-free and few-shot conditions, named as SF-FSDA. To overcome this problem, we develop an efficient labeled data factory based approach. Without accessing the source domain, the data factory renders i) infinite amount of synthesized target-domain like images, under the guidance of the few-shot image samples and text description from the target domain; ii) corresponding bounding box and category annotations, only demanding minimum human effort, i.e., a few manually labeled examples. On the one hand, the synthesized images mitigate the knowledge insufficiency brought by the few-shot condition. On the other hand, compared to the popular pseudo-label technique, the generated annotations from data factory not only get rid of the reliance on the source pretrained object detection model, but also alleviate the unavoidably pseudo-label noise due to domain shift and source-free condition. The generated dataset is further utilized to adapt the source pretrained object detection model, realizing the robust object detection under SF-FSDA. The experiments on different settings showcase that our proposed approach outperforms other state-of-the-art methods on SF-FSDA problem. Our codes and models will be made publicly available.' volume: 232 URL: https://proceedings.mlr.press/v232/sun23a.html PDF: https://proceedings.mlr.press/v232/sun23a/sun23a.pdf edit: https://github.com/mlresearch//v232/edit/gh-pages/_posts/2023-11-20-sun23a.md series: 'Proceedings of Machine Learning Research' container-title: 'Proceedings of The 2nd Conference on Lifelong Learning Agents' publisher: 'PMLR' author: - given: Han family: Sun - given: Rui family: Gong - given: Konrad family: Schindler - given: Luc family: Van Gool editor: - given: Sarath family: Chandar - given: Razvan family: Pascanu - given: Hanie family: Sedghi - given: Doina family: Precup page: 1091-1111 id: sun23a issued: date-parts: - 2023 - 11 - 20 firstpage: 1091 lastpage: 1111 published: 2023-11-20 00:00:00 +0000