The HSIC Bottleneck: Deep Learning without Back-Propagation

Authors: Wan-Duo Kurt Ma, J. P. Lewis, W. Bastiaan Kleijn5085-5092

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We find that the HSIC bottleneck provides performance on MNIST/Fashion MNIST/CIFAR10 classification comparable to backpropagation with a cross-entropy target, even when the system is not encouraged to make the output resemble the classification labels. In this section, we report several experiments that explore and validate the HSIC-trained network concept.
Researcher Affiliation Collaboration Victoria University {mawand, bastiaan.kleijn}@ecs.vuw.ac.nz, jplewis@google.com
Pseudocode Yes Algorithm 1: Unformatted-training
Open Source Code Yes Our code is available at https://github.com/choasma/HSIC-Bottleneck
Open Datasets Yes For the experiments, we used standard feedforward networks with batch-normalization (Ioffe and Szegedy 2015) on the MNIST/Fashion MNIST/CIFAR10 datasets.
Dataset Splits No The paper does not provide specific training/validation/test dataset splits, only mentioning 'Typically the minibatch size is a constant that is chosen based on validation performance'.
Hardware Specification No The paper mentions 'available GPU memory' and 'on a GPU' in the context of HSIC complexity, but does not provide specific hardware details such as GPU models, CPU types, or other computer specifications used for running experiments.
Software Dependencies No The paper mentions using 'standard feedforward networks with batch-normalization' and a 'simple SGD optimizer', but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, or CUDA versions).
Experiment Setup Yes All experiments including standard backpropagation, unformattedtrained, and format-trained, use a simple SGD optimizer. The coefficient β and the kernel scale factor σ of the HSIC-bottleneck were set to 500 and 5 respectively, which empirically balances compression and the relevant information available for the classification task. We use a fully connected network architecture 784-256-256-256-256-256-10 with Re LU activation functions.