Training Quantized Neural Networks to Global Optimality via Semidefinite Programming

Authors: Burak Bartan, Mert Pilanci

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present numerical examples to illustrate the effectiveness of our method.
Researcher Affiliation Academia Department of Electrical Engineering, Stanford University, CA, USA.
Pseudocode Yes Algorithm 1 Sampling algorithm for quantized neural networks
Open Source Code No The paper does not provide a direct link or explicit statement about releasing its own source code for the described methodology. It mentions using third-party tools like CVXPY, SCS, and PyTorch.
Open Datasets Yes The dataset is the binary classification breast-cancer dataset and has n = 228 training samples and 58 test samples and the samples are d = 9 dimensional. Figure 2 shows the classification accuracy against time for various methods which we describe below. The regularization coefficient β is picked for each method separately by searching the value that yields the highest accuracy and the resulting β values are provided in the captions of the figures. [...] Figure 3 shows results for the UCI repository dataset ionosphere . This is a binary classification dataset with n = 280 training samples and 71 test samples. The samples are d = 33 dimensional. The experiment setting is similar to Figure 2 with the main difference that the number of neurons is 10 times higher (i.e., m = 2500).
Dataset Splits No The paper specifies training and test sample counts for the datasets used but does not mention a distinct validation split.
Hardware Specification Yes The experiments have been carried out on a Mac Book with 2.2 GHz 6-Core Intel Core i7 processor and 16 GB of RAM.
Software Dependencies No The paper mentions software like CVXPY, SCS, and PyTorch, but does not provide specific version numbers for these dependencies within the text.
Experiment Setup Yes The regularization coefficient is β = 10 4. [...] We fix the second layer weights to 1/m during training. [...] The number of neurons is m = 250 and the regularization coefficient is β = 0.1 for the SDP based method and β = 0.1 for the backpropagation. [...] The number of neurons is 10 times higher (i.e., m = 2500) and the regularization coefficient is β = 10 for the SDP based method, β = 10 6 for backpropagation.