Matching Learned Causal Effects of Neural Networks with Domain Priors

Authors: Sai Srinivas Kancheti, Abbavaram Gowtham Reddy, Vineeth N Balasubramanian, Amit Sharma

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on twelve benchmark datasets show its utility in regularizing a neural network model to maintain desired causal effects, without compromising on accuracy. Importantly, we also show that a model thus trained is robust and gets improved accuracy on noisy inputs.
Researcher Affiliation Collaboration 1Indian Institute of Technology Hyderabad, India 2Microsoft Research, Bangalore, India.
Pseudocode Yes Algorithm 1 CREDO Regularizer and Algorithm 2 Prior function parameter search
Open Source Code Yes Our code is made available for reproducibility.
Open Datasets Yes We use two kinds of datasets: 1) Four benchmark synthetic datasets from the BNLearn repository (Scutari & Denis, 2014)...; and 2) Eight realworld datasets without knowledge of true causal graph... E.g., in the Boston Housing dataset... Auto MPG... COMPAS: The task herein (Angwin et al., 2016)... from the UCI repository (Dua & Graff, 2017).
Dataset Splits Yes 80% of the dataset is used for training and remaining 20% for testing. ... We find the best parameters for the prior function by tuning for highest validation-set accuracy.
Hardware Specification Yes All experiments were conducted on one NVIDIA Ge Force 1080Ti GPU.
Software Dependencies No The paper mentions software components like 'ADAM optimizer' and 'multi-layer perceptron with Re LU non-linearity' but does not specify version numbers for these or other software dependencies.
Experiment Setup Yes We use a multi-layer perceptron with Re LU non-linearity across our experiments, each trained using ADAM optimizer with Dropout and 𝐿2 weight decay. Table 16 shows the details of neural network architectures and training details of our models for various datasets. ... Table 16: S.No. Dataset Input Size, Output Size Learning Rate Batch Size 𝜆1 Number of Layers Size of Each Layer