NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks

Mustafa B Gurbuz, Constantine Dovrolis
Proceedings of the 39th International Conference on Machine Learning, PMLR 162:8157-8174, 2022.

Abstract

The goal of continual learning (CL) is to learn different tasks over time. The main desiderata associated with CL are to maintain performance on older tasks, leverage the latter to improve learning of future tasks, and to introduce minimal overhead in the training process (for instance, to not require a growing model or retraining). We propose the Neuro-Inspired Stability-Plasticity Adaptation (NISPA) architecture that addresses these desiderata through a sparse neural network with fixed density. NISPA forms stable paths to preserve learned knowledge from older tasks. Also, NISPA uses connection rewiring to create new plastic paths that reuse existing knowledge on novel tasks. Our extensive evaluation on EMNIST, FashionMNIST, CIFAR10, and CIFAR100 datasets shows that NISPA significantly outperforms representative state-of-the-art continual learning baselines, and it uses up to ten times fewer learnable parameters compared to baselines. We also make the case that sparsity is an essential ingredient for continual learning. The NISPA code is available at https://github.com/BurakGurbuz97/NISPA.

Cite this Paper


BibTeX
@InProceedings{pmlr-v162-gurbuz22a, title = {{NISPA}: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks}, author = {Gurbuz, Mustafa B and Dovrolis, Constantine}, booktitle = {Proceedings of the 39th International Conference on Machine Learning}, pages = {8157--8174}, year = {2022}, editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan}, volume = {162}, series = {Proceedings of Machine Learning Research}, month = {17--23 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v162/gurbuz22a/gurbuz22a.pdf}, url = {https://proceedings.mlr.press/v162/gurbuz22a.html}, abstract = {The goal of continual learning (CL) is to learn different tasks over time. The main desiderata associated with CL are to maintain performance on older tasks, leverage the latter to improve learning of future tasks, and to introduce minimal overhead in the training process (for instance, to not require a growing model or retraining). We propose the Neuro-Inspired Stability-Plasticity Adaptation (NISPA) architecture that addresses these desiderata through a sparse neural network with fixed density. NISPA forms stable paths to preserve learned knowledge from older tasks. Also, NISPA uses connection rewiring to create new plastic paths that reuse existing knowledge on novel tasks. Our extensive evaluation on EMNIST, FashionMNIST, CIFAR10, and CIFAR100 datasets shows that NISPA significantly outperforms representative state-of-the-art continual learning baselines, and it uses up to ten times fewer learnable parameters compared to baselines. We also make the case that sparsity is an essential ingredient for continual learning. The NISPA code is available at https://github.com/BurakGurbuz97/NISPA.} }
Endnote
%0 Conference Paper %T NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks %A Mustafa B Gurbuz %A Constantine Dovrolis %B Proceedings of the 39th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2022 %E Kamalika Chaudhuri %E Stefanie Jegelka %E Le Song %E Csaba Szepesvari %E Gang Niu %E Sivan Sabato %F pmlr-v162-gurbuz22a %I PMLR %P 8157--8174 %U https://proceedings.mlr.press/v162/gurbuz22a.html %V 162 %X The goal of continual learning (CL) is to learn different tasks over time. The main desiderata associated with CL are to maintain performance on older tasks, leverage the latter to improve learning of future tasks, and to introduce minimal overhead in the training process (for instance, to not require a growing model or retraining). We propose the Neuro-Inspired Stability-Plasticity Adaptation (NISPA) architecture that addresses these desiderata through a sparse neural network with fixed density. NISPA forms stable paths to preserve learned knowledge from older tasks. Also, NISPA uses connection rewiring to create new plastic paths that reuse existing knowledge on novel tasks. Our extensive evaluation on EMNIST, FashionMNIST, CIFAR10, and CIFAR100 datasets shows that NISPA significantly outperforms representative state-of-the-art continual learning baselines, and it uses up to ten times fewer learnable parameters compared to baselines. We also make the case that sparsity is an essential ingredient for continual learning. The NISPA code is available at https://github.com/BurakGurbuz97/NISPA.
APA
Gurbuz, M.B. & Dovrolis, C.. (2022). NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks. Proceedings of the 39th International Conference on Machine Learning, in Proceedings of Machine Learning Research 162:8157-8174 Available from https://proceedings.mlr.press/v162/gurbuz22a.html.

Related Material