Towards Scaling Difference Target Propagation by Learning Backprop Targets

Maxence M Ernoult, Fabrice Normandin, Abhinav Moudgil, Sean Spinney, Eugene Belilovsky, Irina Rish, Blake Richards, Yoshua Bengio
Proceedings of the 39th International Conference on Machine Learning, PMLR 162:5968-5987, 2022.

Abstract

The development of biologically-plausible learning algorithms is important for understanding learning in the brain, but most of them fail to scale-up to real-world tasks, limiting their potential as explanations for learning by real brains. As such, it is important to explore learning algorithms that come with strong theoretical guarantees and can match the performance of backpropagation (BP) on complex tasks. One such algorithm is Difference Target Propagation (DTP), a biologically-plausible learning algorithm whose close relation with Gauss-Newton (GN) optimization has been recently established. However, the conditions under which this connection rigorously holds preclude layer-wise training of the feedback pathway synaptic weights (which is more biologically plausible). Moreover, good alignment between DTP weight updates and loss gradients is only loosely guaranteed and under very specific conditions for the architecture being trained. In this paper, we propose a novel feedback weight training scheme that ensures both that DTP approximates BP and that layer-wise feedback weight training can be restored without sacrificing any theoretical guarantees. Our theory is corroborated by experimental results and we report the best performance ever achieved by DTP on CIFAR-10 and ImageNet 32x32.

Cite this Paper

BibTeX
@InProceedings{pmlr-v162-ernoult22a, title = {Towards Scaling Difference Target Propagation by Learning Backprop Targets}, author = {Ernoult, Maxence M and Normandin, Fabrice and Moudgil, Abhinav and Spinney, Sean and Belilovsky, Eugene and Rish, Irina and Richards, Blake and Bengio, Yoshua}, booktitle = {Proceedings of the 39th International Conference on Machine Learning}, pages = {5968--5987}, 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/ernoult22a/ernoult22a.pdf}, url = {https://proceedings.mlr.press/v162/ernoult22a.html}, abstract = {The development of biologically-plausible learning algorithms is important for understanding learning in the brain, but most of them fail to scale-up to real-world tasks, limiting their potential as explanations for learning by real brains. As such, it is important to explore learning algorithms that come with strong theoretical guarantees and can match the performance of backpropagation (BP) on complex tasks. One such algorithm is Difference Target Propagation (DTP), a biologically-plausible learning algorithm whose close relation with Gauss-Newton (GN) optimization has been recently established. However, the conditions under which this connection rigorously holds preclude layer-wise training of the feedback pathway synaptic weights (which is more biologically plausible). Moreover, good alignment between DTP weight updates and loss gradients is only loosely guaranteed and under very specific conditions for the architecture being trained. In this paper, we propose a novel feedback weight training scheme that ensures both that DTP approximates BP and that layer-wise feedback weight training can be restored without sacrificing any theoretical guarantees. Our theory is corroborated by experimental results and we report the best performance ever achieved by DTP on CIFAR-10 and ImageNet 32x32.} }
Endnote
%0 Conference Paper %T Towards Scaling Difference Target Propagation by Learning Backprop Targets %A Maxence M Ernoult %A Fabrice Normandin %A Abhinav Moudgil %A Sean Spinney %A Eugene Belilovsky %A Irina Rish %A Blake Richards %A Yoshua Bengio %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-ernoult22a %I PMLR %P 5968--5987 %U https://proceedings.mlr.press/v162/ernoult22a.html %V 162 %X The development of biologically-plausible learning algorithms is important for understanding learning in the brain, but most of them fail to scale-up to real-world tasks, limiting their potential as explanations for learning by real brains. As such, it is important to explore learning algorithms that come with strong theoretical guarantees and can match the performance of backpropagation (BP) on complex tasks. One such algorithm is Difference Target Propagation (DTP), a biologically-plausible learning algorithm whose close relation with Gauss-Newton (GN) optimization has been recently established. However, the conditions under which this connection rigorously holds preclude layer-wise training of the feedback pathway synaptic weights (which is more biologically plausible). Moreover, good alignment between DTP weight updates and loss gradients is only loosely guaranteed and under very specific conditions for the architecture being trained. In this paper, we propose a novel feedback weight training scheme that ensures both that DTP approximates BP and that layer-wise feedback weight training can be restored without sacrificing any theoretical guarantees. Our theory is corroborated by experimental results and we report the best performance ever achieved by DTP on CIFAR-10 and ImageNet 32x32.
APA
Ernoult, M.M., Normandin, F., Moudgil, A., Spinney, S., Belilovsky, E., Rish, I., Richards, B. & Bengio, Y.. (2022). Towards Scaling Difference Target Propagation by Learning Backprop Targets. Proceedings of the 39th International Conference on Machine Learning, in Proceedings of Machine Learning Research 162:5968-5987 Available from https://proceedings.mlr.press/v162/ernoult22a.html.

Related Material