Energy-based learning algorithms for analog computing: a comparative study

Authors: Benjamin Scellier, Maxence Ernoult, Jack Kendall, Suhas Kumar

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we carry out a comparison of seven learning algorithms, namely CL and different variants of EP and Cp L depending on the signs of the perturbations. Specifically, using these learning algorithms, we train deep convolutional Hopfield networks (DCHNs) on five vision tasks (MNIST, F-MNIST, SVHN, CIFAR-10 and CIFAR-100). We find that, while all algorithms yield comparable performance on MNIST, important differences in performance arise as the difficulty of the task increases.
Researcher Affiliation Industry Benjamin Scellier Rain AI benjamin@rain.ai Maxence Ernoult Rain AI maxence@rain.ai Jack Kendall Rain AI jack@rain.ai Suhas Kumar Rain AI suhas@rain.ai
Pseudocode No The paper describes the energy-based learning algorithms and procedures in narrative text and mathematical equations, but it does not contain a dedicated pseudocode or algorithm block.
Open Source Code Yes The code is available at https://github.com/rain-neuromorphics/energy-based-learning
Open Datasets Yes We train a DCHN on MNIST, Fashion-MNIST, SVHN, CIFAR-10 and CIFAR-100 using each of the seven EBL algorithms. [...] The MNIST dataset is composed of images of handwritten digits [Le Cun et al., 1998]. [...] The CIFAR-10 dataset [Krizhevsky, Hinton, et al., 2009] consists of 60,000 colour images of 32x32 pixels.
Dataset Splits No The paper details training and test set sizes for each dataset (e.g., "The dataset contains 60,000 training images and 10,000 test images" for MNIST, and "The training set consists of 50,000 images and the test set of 10,000 images" for CIFAR-10), but does not explicitly provide information on validation dataset splits or how a validation set was used.
Hardware Specification Yes Each run is performed on a single A100 GPU. [...] For example, a full run of 100 epochs on CIFAR-10 yields a test error rate of 10.4% and takes 3 hours 18 minutes on a single A100 GPU
Software Dependencies Yes The code for the simulations uses PyTorch 1.13.1 and TorchVision 0.14.1. Paszke et al. [2017].
Experiment Setup Yes Table 5 contains the architectural details of the network, as well as the hyperparameters used to obtain the results presented in Table 1, Table 2 and Table 6. [...] nudging (β) 0.25 num. iterations at inference (T) 60 num. iterations at training (K) 15 [...] lr conv-weight 1 & bias 1 (η1) 0.0625 momentum 0.9 weight decay (1e-4) 3.0 mini-batch size 128 number of epochs 100 Tmax 100 ηmin (1e-6) 2.0