How Important is a Neuron

Authors: Kedar Dhamdhere, Mukund Sundararajan, Qiqi Yan

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We justify conductance in multiple ways via a qualitative comparison with other methods, via some axiomatic results, and via an empirical evaluation based on a feature selection task. The empirical evaluations are done using the Inception network over Image Net data, and a convolutional network over text data. In both cases, we demonstrate the effectiveness of conductance in identifying interesting insights about the internal workings of these networks.
Researcher Affiliation Industry Kedar Dhamdhere Google AI kedar@google.com Mukund Sundararajan Google AI mukunds@google.com Qiqi Yan Google AI qiqiyan@google.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statements or links regarding the open-sourcing of the code for the methodology described.
Open Datasets Yes The empirical evaluations are done using the Inception network over Image Net data (Russakovsky et al. (2015)), and a convolutional network over text data. ... We analyze a CNN model ... We consider the Wiki Table Questions data set introduced by Pasupat & Liang (2015).
Dataset Splits Yes This dataset has 1000 labels with about 1000 training and 50 validation images per label. ... We picked about 50 images per label from the validation set we split this into 30 images for training and 20 for eval.
Hardware Specification No The paper does not specify any particular hardware (e.g., CPU, GPU models, or cloud computing instance types) used for running the experiments.
Software Dependencies No The paper mentions machine-learning frameworks like 'tensorflow or pytorch' generally but does not list any specific software dependencies with version numbers.
Experiment Setup Yes The network is built using the Google Net architecture Szegedy et al. (2014) and trained over the Image Net object recognition dataset Russakovsky et al. (2015). ... The model is a convolutional model from Kim (2014b) trained over review data. In the variant of the model we study, the words are embedded into a 50 dimensional space. The embeddings are fed to a convolutional layer that has 4 window sizes ranging from 3 to 6 words. Each of the four filter widths has 64 feature maps. Each of the 4 * 64 = 256 filters is 1-max pooled to construct a layer of size 256, which is then fed (fully connected) to a layer of 10 neurons, which is fed to a logit, that feeds into a softmax that classifies whether the sentiment is positive and/or negative (two binary heads).