Credit Assignment in Neural Networks through Deep Feedback Control

Authors: Alexander Meulemans, Matilde Tristany Farinha, Javier Garcia Ordonez, Pau Vilimelis Aceituno, João Sacramento, Benjamin F. Grewe

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we provide detailed experimental results, corroborating our theoretical contributions and showing that DFC does principled CA on standard computer-vision benchmarks in a way that fundamentally differs from standard BP. We evaluate DFC in detail on toy experiments to showcase that our theoretical results translate to practice (Section 6.1) and on a modest range of computer vision benchmarks MNIST classification and autoencoding [40], and Fashion MNIST classification [41] to show that DFC can do precise CA in more challenging settings (Section 6.2).
Researcher Affiliation Academia Institute of Neuroinformatics, University of Zürich and ETH Zürich ameulema@ethz.ch
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes PyTorch implementation of all methods is available at https://github.com/meulemansalex/deep_feedback_control.
Open Datasets Yes We evaluate DFC in detail on toy experiments to showcase that our theoretical results translate to practice (Section 6.1) and on a modest range of computer vision benchmarks MNIST classification and autoencoding [40], and Fashion MNIST classification [41] to show that DFC can do precise CA in more challenging settings (Section 6.2).
Dataset Splits Yes Test errors (classification) and test loss (autoencoder) corresponding to the epoch with the best validation result (for 5000 validation samples) over a training of 100 epochs (classification) or 25 epochs (autoencoder).
Hardware Specification No The paper discusses the potential for DFC implementation on analog hardware in the future but does not provide any specific details (e.g., GPU/CPU models, memory amounts) about the hardware used to run the experiments reported in the paper.
Software Dependencies No The paper mentions 'PyTorch implementation' and 'Adam optimizer' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Architectures: 3x256 fully connected (FC) tanh hidden layers and softmax output (classification), 256-32-256 FC hidden layers for autoencoder MNIST with tanh-linear-tanh nonlinearities, and a linear output. Results for nonlinear student-teacher regression task with layer sizes (15-10-10-5), tanh nonlinearities, a linear output layer, kp = 1.5, λ = 0.05, and α = 0.0015.