Neuron Shapley: Discovering the Responsible Neurons

Authors: Amirata Ghorbani, James Y. Zou

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical: our systematic experiments discover several interesting findings, including the phenomenon that a small number of neurons are critical to different aspects of a network s performance, e.g. accuracy, fairness, robustness. This facilitates both model interpretation and repair.
Researcher Affiliation Academia Amirata Ghorbani Department of Electrical Engineering Stanford University Stanford, CA 94025 amiratag@stanford.edu; James Zou Department of Biomedical Data Science Stanford University Stanford, CA 94025 jamesz@stanford.edu
Pseudocode Yes Algorithm 1 Truncated Multi Armed Bandit Shapley
Open Source Code Yes Code is available on Github at https://github.com/amiratag/neuronshapley
Open Datasets Yes First is the Inception-v3 [44] architecture trained on the ILSVRC2012 (a.k.a Image Net) [36] dataset... The second model is the Squeeze Net [16] architecture... trained on the celeb A [25] dataset
Dataset Splits Yes We divide the released Image Net validation set into two parts (25000 images each) to serve as validation and test sets.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types) used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup Yes We set k = 100 to detect the top-100 important filters. We run the original MC-Shapley algorithm and our TMAB-Shapley algorithm on the Squeezenet model (δ = 0.05, = 10 4). We use iterative PGD attack [22, 28] as an adversary... we use 1 perturbations with size = 16/255.