Multiple Instance Learning with Absolute Position Information

Meera Krishnamoorthy, Jenna Wiens
Proceedings of the fifth Conference on Health, Inference, and Learning, PMLR 248:88-104, 2024.

Abstract

Most past work in multiple instance learning (MIL), which maps a group or bag of instances to a classification label, has focused on settings in which the order of instances does not contain information. In this paper, we define MIL with \textit{absolute} position information: tasks in which instances of importance remain in similar positions across bags. Such problems arise, for example, in MIL with medical images in which there exists a common global alignment across images (e.g., in chest x-rays the heart is in a similar location). We also evaluate the performance of existing MIL methods on a set of new benchmark tasks and two real data tasks with varying amounts of absolute position information. We find that, despite being less computationally efficient than other approaches, transformer-based MIL methods are more accurate at classifying tasks with absolute position information. Thus, we investigate the ability of positional encodings, a mechanism typically only used in transformers, to improve the accuracy of other MIL approaches. Applied to the task of identifying pathological findings in chest x-rays, when augmented with positional encodings, standard MIL approaches perform significantly better than without (AUROC of 0.799, 95% CI: [0.791, 0.806] vs. 0.782, 95% CI: [0.774, 0.789]) and on-par with transformer-based methods (AUROC of 0.797, 95% CI: [0.790, 0.804]) while being 10 times faster. Our results suggest that one can efficiently and accurately classify MIL data with absolute position information using standard approaches by simply including positional encodings.

Cite this Paper

BibTeX
@InProceedings{pmlr-v248-krishnamoorthy24a, title = {Multiple Instance Learning with Absolute Position Information}, author = {Krishnamoorthy, Meera and Wiens, Jenna}, booktitle = {Proceedings of the fifth Conference on Health, Inference, and Learning}, pages = {88--104}, year = {2024}, editor = {Pollard, Tom and Choi, Edward and Singhal, Pankhuri and Hughes, Michael and Sizikova, Elena and Mortazavi, Bobak and Chen, Irene and Wang, Fei and Sarker, Tasmie and McDermott, Matthew and Ghassemi, Marzyeh}, volume = {248}, series = {Proceedings of Machine Learning Research}, month = {27--28 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v248/main/assets/krishnamoorthy24a/krishnamoorthy24a.pdf}, url = {https://proceedings.mlr.press/v248/krishnamoorthy24a.html}, abstract = {Most past work in multiple instance learning (MIL), which maps a group or bag of instances to a classification label, has focused on settings in which the order of instances does not contain information. In this paper, we define MIL with \textit{absolute} position information: tasks in which instances of importance remain in similar positions across bags. Such problems arise, for example, in MIL with medical images in which there exists a common global alignment across images (e.g., in chest x-rays the heart is in a similar location). We also evaluate the performance of existing MIL methods on a set of new benchmark tasks and two real data tasks with varying amounts of absolute position information. We find that, despite being less computationally efficient than other approaches, transformer-based MIL methods are more accurate at classifying tasks with absolute position information. Thus, we investigate the ability of positional encodings, a mechanism typically only used in transformers, to improve the accuracy of other MIL approaches. Applied to the task of identifying pathological findings in chest x-rays, when augmented with positional encodings, standard MIL approaches perform significantly better than without (AUROC of 0.799, 95% CI: [0.791, 0.806] vs. 0.782, 95% CI: [0.774, 0.789]) and on-par with transformer-based methods (AUROC of 0.797, 95% CI: [0.790, 0.804]) while being 10 times faster. Our results suggest that one can efficiently and accurately classify MIL data with absolute position information using standard approaches by simply including positional encodings.} }
Endnote
%0 Conference Paper %T Multiple Instance Learning with Absolute Position Information %A Meera Krishnamoorthy %A Jenna Wiens %B Proceedings of the fifth Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2024 %E Tom Pollard %E Edward Choi %E Pankhuri Singhal %E Michael Hughes %E Elena Sizikova %E Bobak Mortazavi %E Irene Chen %E Fei Wang %E Tasmie Sarker %E Matthew McDermott %E Marzyeh Ghassemi %F pmlr-v248-krishnamoorthy24a %I PMLR %P 88--104 %U https://proceedings.mlr.press/v248/krishnamoorthy24a.html %V 248 %X Most past work in multiple instance learning (MIL), which maps a group or bag of instances to a classification label, has focused on settings in which the order of instances does not contain information. In this paper, we define MIL with \textit{absolute} position information: tasks in which instances of importance remain in similar positions across bags. Such problems arise, for example, in MIL with medical images in which there exists a common global alignment across images (e.g., in chest x-rays the heart is in a similar location). We also evaluate the performance of existing MIL methods on a set of new benchmark tasks and two real data tasks with varying amounts of absolute position information. We find that, despite being less computationally efficient than other approaches, transformer-based MIL methods are more accurate at classifying tasks with absolute position information. Thus, we investigate the ability of positional encodings, a mechanism typically only used in transformers, to improve the accuracy of other MIL approaches. Applied to the task of identifying pathological findings in chest x-rays, when augmented with positional encodings, standard MIL approaches perform significantly better than without (AUROC of 0.799, 95% CI: [0.791, 0.806] vs. 0.782, 95% CI: [0.774, 0.789]) and on-par with transformer-based methods (AUROC of 0.797, 95% CI: [0.790, 0.804]) while being 10 times faster. Our results suggest that one can efficiently and accurately classify MIL data with absolute position information using standard approaches by simply including positional encodings.
APA
Krishnamoorthy, M. & Wiens, J.. (2024). Multiple Instance Learning with Absolute Position Information. Proceedings of the fifth Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 248:88-104 Available from https://proceedings.mlr.press/v248/krishnamoorthy24a.html.

Related Material