From Boltzmann Machines to Neural Networks and Back Again

Authors: Surbhi Goel, Adam Klivans, Frederic Koehler

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

Reproducibility Variable Result LLM Response
Research Type Experimental Lastly, we present an experimental evaluation of our "supervised RBM" algorithm on MNIST and Fashion MNIST to highlight the applicability of our techniques in practice (Section 5). In this section we present some simple experiments on MNIST and Fashion MNIST to confirm that our method performs reasonably well in practice.
Researcher Affiliation Collaboration Surbhi Goel Microsoft Research NYC surbgoel@microsoft.com Adam Klivans University of Texas at Austin klivans@cs.utexas.edu Frederic Koehler MIT fkoehler@mit.edu
Pseudocode Yes Algorithm 1 LEARNSUPERVISEDRBMNBHD(u, τ, S) (Adapted from [18, 19])
Open Source Code No The paper does not contain an explicit statement or link providing access to open-source code for the described methodology.
Open Datasets Yes Lastly, we present an experimental evaluation of our "supervised RBM" algorithm on MNIST and Fashion MNIST to highlight the applicability of our techniques in practice (Section 5). For Fashion MNIST, we obtained a test accuracy of 88.84 0.31%; the training accuracy was 92.19% and we trained the logistic regression for 45 epochs with L-BFGS as before. Both datasets have 60,000 training points and 10,000 test;
Dataset Splits No The paper states: "Both datasets have 60,000 training points and 10,000 test;" but does not specify a validation set split.
Hardware Specification Yes Overall training took a bit less than an hour each on a Kaggle notebook with a P100 GPU.
Software Dependencies No The paper mentions "logistic regression", "L-BFGS", and "Gibbs sampling" but does not specify software names with version numbers (e.g., Python, PyTorch, TensorFlow, or scikit-learn versions).
Experiment Setup Yes we trained the logistic regression for 30 epochs (same as steps) of L-BFGS with line search enabled. For Fashion MNIST, we obtained a test accuracy of 88.84 0.31%; the training accuracy was 92.19% and we trained the logistic regression for 45 epochs with L-BFGS as before. Overall training took a bit less than an hour each on a Kaggle notebook with a P100 GPU. in both experiments we used a maximum neighborhood size of 12, and stopped adding neighbors if the conditional variance shrunk by less than 1%.