Debiasing Concept-based Explanations with Causal Analysis

Authors: Mohammad Taha Bahadori, David Heckerman

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our synthetic and real-world experiments demonstrate the success of our method in removing biases and improving the ranking of the concepts in terms of their contribution to the explanation of the predictions.
Researcher Affiliation Industry Mohammad Taha Bahadori, David E. Heckerman {bahadorm, heckerma}@amazon.com
Pseudocode Yes Algorithm 1 Debiased CBMs
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of the described methodology.
Open Datasets Yes We evaluate the performance of the proposed approach on the CUB200-2011 dataset (Wah et al., 2011).
Dataset Splits Yes The dataset includes 11788 pictures (in 5994/5794 train/test partitions) of 200 different types of birds, annotated both for the bird type and 312 different concepts about each picture. ... We randomly choose 15% of the training set and hold out as the validation set.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running experiments.
Software Dependencies No The paper mentions 'Torch Vision s pre-trained Res Net152 network' and 'Adam optimization algorithm' but does not specify their version numbers or other crucial software dependencies with versions.
Experiment Setup Yes We model the distribution of the concept logits as Gaussians with means equal to the Res Net152 s logit outputs and a diagonal covariance matrix. We estimate the variance for each concept by using the logits of the true concept annotation scores that are clamped into [0.05, 0.95] to avoid large logit numbers. In each iteration of the training loop for Line 3, we draw 25 samples from the estimated p(d|x). Predictor of labels from concepts (the function g( ) in Eq. (5)) is a three-layer feed-forward neural network with hidden layer sizes (312, 312, 200). There is a skip connection from the input to the penultimate layer. All algorithms are trained with Adam optimization algorithm (Kingma & Ba, 2014).