Neuron Dependency Graphs: A Causal Abstraction of Neural Networks

Authors: Yaojie Hu, Jin Tian

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that an NDG is a causal abstraction of the corresponding neural network that unfolds the same way under causal interventions using the theory by Geiger et al. (2021a). Code is available at https://github.com/phimachine/ndg. ... Experimental results empirically support the alignment condition. ... We extract neuron dependency graphs on a diverse set of datasets and architectures to demonstrate the generality of our method. Table 1 lists the datasets and architectures
Researcher Affiliation Academia Yaojie Hu 1 Jin Tian 1 ... 1Department of Computer Science, Iowa State University, United States. Correspondence to: Yaojie Hu <jhu@iastate.edu>.
Pseudocode Yes Algorithm 1 Interchange intervention with a NDG
Open Source Code Yes Code is available at https://github.com/phimachine/ndg.
Open Datasets Yes We extract neuron dependency graphs on a diverse set of datasets and architectures to demonstrate the generality of our method. Table 1 lists the datasets and architectures (Le Cun et al., 1998; Socher et al., 2013; Conneau et al., 2017; Reimers & Gurevych, 2019; Welinder et al., 2010; Lu et al., 2021; Sanh et al., 2019; Dosovitskiy et al., 2021; He et al., 2021; Zhou et al., 2019; Feng et al., 2020; Liu et al., 2019; Devlin et al., 2018).
Dataset Splits Yes For the datasets with only the training set and the test set, we leave 10% of the training set for validation.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, or cloud instance types used for experiments.
Software Dependencies No The Py Torch program using Huggingface library (Wolf et al., 2019) to select layers is in Figure 4.
Experiment Setup Yes Threshold parameter α is used in Eq. (1) to extract the neuron dependency graphs. ... Threshold α is manually selected to improve interchange intervention accuracy. ... For Re LU, we select Tϕ = 1, Fϕ = 0. For sigmoid, Tϕ = , Fϕ = .