Convergence and Alignment of Gradient Descent with Random Backpropagation Weights

Authors: Ganlin Song, Ruitu Xu, John Lafferty

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

Reproducibility Variable Result LLM Response
Research Type Experimental In the overparameterized setting, we prove that the error converges to zero exponentially fast, and also that regularization is necessary in order for the parameters to become aligned with the random backpropagation weights. Simulations are given that are consistent with this analysis and suggest further generalizations.
Researcher Affiliation Academia Ganlin Song Ruitu Xu John Lafferty Department of Statistics and Data Science Wu Tsai Institute Yale University {ganlin.song, ruitu.xu, john.lafferty}@yale.edu
Pseudocode Yes Algorithm 1 Feedback Alignment Input: Dataset {(xi, yi)}n i=1, step size η 1: initialize W, β and b as Gaussian 2: while not converged do 3: βr βr η p Pn i=1 eiψ(w r xi) 4: wr wr η p Pn i=1 eibrψ (w r xi)xi 5: for r [p] 6: end while
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes The MNIST dataset is available under the Creative Commons Attribution-Share Alike 3.0 license (Deng, 2012).
Dataset Splits Yes It consists of 60,000 training images and 10,000 test images of dimension 28 by 28.
Hardware Specification Yes We implement the feedback alignment procedure in Py Torch as an extension of the autograd module for backpropagation, and the training is done on V100 GPUs from internal clusters.
Software Dependencies No The paper mentions 'Py Torch' but does not provide a specific version number for the software dependency.
Experiment Setup Yes During training, we take step size η = 10 4 for linear networks and η = 10 3, 10 2 for Re LU and Tanh networks, respectively.