Deep Learning without Weight Transport

Authors: Mohamed Akrout, Collin Wilson, Peter Humphreys, Timothy Lillicrap, Douglas B. Tweed

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Tested on the Image Net visual-recognition task, these mechanisms learn almost as well as backprop (the standard algorithm of deep learning, which uses weight transport) and they outperform feedback alignment and another, more-recent transport-free algorithm, the sign-symmetry method.
Researcher Affiliation Collaboration Mohamed Akrout University of Toronto, Triage Collin Wilson University of Toronto Peter C. Humphreys Deep Mind Timothy Lillicrap Deep Mind, University College London Douglas Tweed University of Toronto, York University
Pseudocode No The paper provides equations and diagrams of neural circuits but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes We have released a Python version of the proprietary Tensor Flow/TPU code for the weight mirror and the KP reciprocal network that we used in our tests; see github.com/makrout/Deep-Learning-without-Weight-Transport.
Open Datasets Yes Tested on the Image Net visual-recognition task [6]
Dataset Splits No The paper mentions ImageNet and discussions of hyperparameter selection in Appendix D.1, implying a standard validation process. However, it does not explicitly state the specific percentages or sample counts for training, validation, or test splits in the main text needed for reproduction.
Hardware Specification Yes We have released a Python version of the proprietary Tensor Flow/TPU code for the weight mirror and the KP reciprocal network that we used in our tests; see github.com/makrout/Deep-Learning-without-Weight-Transport.
Software Dependencies No The paper mentions 'Python version' and 'Tensor Flow/TPU code' but does not specify exact version numbers for these software components.
Experiment Setup No The paper mentions the use of 'Res Net block variant', 'Batch Norm', 'Re LUs', 'learning rate ηW', and 'weight-decay factor λ', and refers to 'Appendix D.1 for details of our hyperparameter selection'. However, it does not provide concrete hyperparameter values or detailed training configurations directly in the main text.