Linear Explanations for Individual Neurons

Authors: Tuomas Oikarinen, Tsui-Wei Weng

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, where we evaluate and compare our method and existing automated explanation methods. and 5. Experiment Results
Researcher Affiliation Academia 1CSE, UC San Diego, CA, USA 2HDSI, UC San Diego, CA, USA. Correspondence to: Tuomas Oikarinen <toikarinen@ucsd.edu>, Tsui-Wei Weng <lweng@ucsd.edu>.
Pseudocode Yes Algorithm 1 Greedy search. Python pseudo-code.
Open Source Code Yes Our code and results are available at https://github.com/TrustworthyML-Lab/Linear-Explanations.
Open Datasets Yes Res Net-50 (Image Net), Res Net-18 (Places365), VGG-16 (CIFAR-100), Vi T-B/16 (Image Net)
Dataset Splits Yes Throughout the paper we use a 70-10-20 split to divide our Dprobe into train, validation and test set splits to avoid overfitting our explanations.
Hardware Specification Yes Both of our explanation methods take around 4 seconds per explained neuron on a machine with a single Tesla V100 GPU.
Software Dependencies No No specific version numbers for key software components like programming languages or deep learning frameworks were provided. It mentions 'GLM-Saga package' and 'Sig LIP' but without versions.
Experiment Setup Yes In our experiments, (v, r, ϵ) are set to be 10, 10, 0.02 respectively. and where Rη(wk) = (1 η) 1 2||wk||2 2 + η||wk||1 and η is a hyperparameter, set to 0.99 in our experiments.