Inner Product-based Neural Network Similarity

Authors: Wei Chen, Zichen Miao, Qiang Qiu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We provide both theoretical and empirical evidence that such simplified filter subspace-based similarity preserves a strong linear correlation with other popular probing-based metrics, while being significantly more efficient to obtain and robust to probing data.
Researcher Affiliation Academia Wei Chen , Zichen Miao , Qiang Qiu Department of ECE Purdue University {chen2732, miaoz, qqiu}@purdue.edu
Pseudocode No The paper describes methods and theoretical proofs but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper states: 'The codes are adapted from https://github.com/lgcollins/Fed Rep' referring to code used for comparison, not the authors' own source code for the proposed method. No direct link or statement is provided for the authors' implementation.
Open Datasets Yes 10-Split CIFAR-100 dataset... CIFAR-100 dataset [23]... Image Net [47]... SVHN [38]
Dataset Splits No The paper mentions distributing data to clients and breaking down datasets into tasks, and uses standard datasets like CIFAR-100 and ImageNet which have standard splits. However, it does not explicitly state the specific training/validation dataset split percentages or sample counts used for experiments.
Hardware Specification Yes For each method, the training takes about 12 hours on Nvidia RTX A5000.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as programming languages, libraries, or frameworks used for implementation.
Experiment Setup Yes All models are trained for T = 100 communication rounds on datasets. At each round, the client executes 1 epoch of SGD with momentum to train the local model, the learning rate is 0.01 and the momentum is 0.9.