Coin Flipping Neural Networks

Authors: Yuval Sieradzki, Nitzan Hodos, Gal Yehuda, Assaf Schuster

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we verify our proofs experimentally using novel CFNN architectures on CIFAR10 and CIFAR100, reaching an improvement of 9.25% from the baseline. In this section we experiment with CFNNs as classifiers of CIFAR10 and CIFAR100 (Krizhevsky, 2009). We present 3 experiments: a study of a hypernetwork architecture inspired by Theorem 4.1; Res Net networks with Dropout viewed as CFNNs; and an analysis of CFNN s accuracy as the number of amplification samples changes. In all experiments, empirical random accuracy was used as the target metric (see Definition 2.1).
Researcher Affiliation Academia Yuval Sieradzki 1 Nitzan Hodos 1 Gal Yehuda 1 Assaf Schuster 1 1Department of Computer Science, Technion Israel Institute of Technology, Haifa, Israel. Correspondence to: Yuval Sieradzki <syuvsier@campus.technion.ac.il>, Nitzan Hodos <hodosnitzan@campus.technion.ac.il>, Assaf Schuster <assaf@technion.ac.il>.
Pseudocode No No clearly labeled 'Pseudocode' or 'Algorithm' blocks were found in the paper. The methodology is described in narrative text and mathematical formulas.
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a code repository for the methodology described.
Open Datasets Yes In this section we experiment with CFNNs as classifiers of CIFAR10 and CIFAR100 (Krizhevsky, 2009).
Dataset Splits Yes In this section we experiment with CFNNs as classifiers of CIFAR10 and CIFAR100 (Krizhevsky, 2009).
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, or detailed specifications of the machines used for experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., libraries, frameworks, or programming languages) used in the experiments.
Experiment Setup Yes Training Hyperparameters Our CFNN is optimized using SGD with Cross Entropy loss, momentum=0.9 and L2 weight decay of 5 10 3, performing majority as described in Section 5. We train our models for 200 epochs with batch size 128. We apply a cosine learning rate scheduler with an initial learning rate 0.1. For CIFAR10, we use n = 25 samples and for CIFAR100 we use n = 125 samples.