Generative causal explanations of black-box classifiers

Authors: Matthew O'Shaughnessy, Gregory Canal, Marissa Connor, Christopher Rozell, Mark Davenport

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we generate explanations of CNN classifiers trained on image recognition tasks, letting G be a set of neural networks and adopting the VAE architecture shown in Figure 1(a) to learn g. Qualitative results. We train a CNN classifier with two convolutional layers followed by two fully connected layers on MNIST 3 and 8 digits...
Researcher Affiliation Academia School of Electrical & Computer Engineering Georgia Institute of Technology
Pseudocode Yes Algorithm 1 Principled procedure for selecting (K, L, λ).
Open Source Code Yes Code is available at https://github.com/siplab-gt/generative-causal-explanations.
Open Datasets Yes We train a CNN classifier with two convolutional layers followed by two fully connected layers on MNIST 3 and 8 digits. We next learn explanations of a CNN trained to classify t-shirt, dress, and coat images from the Fashion MNIST dataset [73].
Dataset Splits Yes For the MNIST experiment, we train on the 3s and 8s digits from the MNIST training set. Validation and test sets are constructed from the remainder of the MNIST training and test sets respectively. We use the Fashion MNIST training set for training, and the Fashion MNIST test set as the validation set for the explanation model (after classifying with the pre-trained black-box classifier).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies or library versions (e.g., PyTorch version, Python version) needed to replicate the experiment.
Experiment Setup Yes Using the parameter tuning procedure described in Algorithm 1, we select K = 1 causal factor, L = 7 noncausal factors, and λ = 0.05. Our black-box classifier is a standard CNN with two convolutional layers followed by two fully connected layers. Both convolutional layers use 3 × 3 kernels, stride 1, and ReLU activations, and are followed by 2 × 2 max pooling layers. The first convolutional layer has 32 filters, and the second has 64 filters. The first fully connected layer has 1024 units. The classifier is trained with Adam optimizer, an initial learning rate of 10−3, a batch size of 64, and a weight decay of 10−4 for 10 epochs. The VAE encoder consists of four convolutional layers, and the decoder is a mirror of the encoder (see Appendix E.4). The VAE is trained for 100 epochs with an initial learning rate of 10−4 and a batch size of 64. The latent dimension of the VAE is 8. The first causal factor is α1 and the remaining seven are noncausal.