Attention-based Interpretability with Concept Transformers

Authors: Mattia Rigotti, Christoph Miksovic, Ioana Giurgiu, Thomas Gschwind, Paolo Scotton

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our Concept Transformer module on established explainability benchmarks and show how it can be used to infuse domain knowledge into classifiers to improve accuracy, and conversely to extract concept-based explanations of classification outputs.
Researcher Affiliation Industry Mattia Rigotti, Christoph Miksovic, Ioana Giurgiu, Thomas Gschwind & Paolo Scotton IBM Research Zurich, Switzerland {mrg,cmi,igi,thg,psc}@zurich.ibm.com
Pseudocode Yes C APPENDIX: CONCEPTTRANSFORMER PYTORCH CODE
Open Source Code Yes Code to reproduce our results is available at: https://github.com/ibm/concept_transformer.
Open Datasets Yes We validate our approach on three image benchmark datasets, MNIST Even/Odd (Barbiero et al., 2021), CUB-200-2011 (Welinder et al., 2010), and a PY (Farhadi et al., 2009).
Dataset Splits Yes Figure 2 shows the accuracy on the test set (left) and explanation loss during validation (right), relative to the number of samples used at training, which varies from 100 to 7000. [...] For validation and testing, only resizing and normalization were applied.
Hardware Specification No The paper does not specify the hardware used for the experiments (e.g., GPU models, CPU types, or memory).
Software Dependencies No We use the Albumentations library by Buslaev et al. (2020). The paper mentions a software library but does not provide specific version numbers for all key software components (e.g., PyTorch version used in Appendix C).
Experiment Setup Yes At training time, the following augmentations were applied to the individual object samples: resizing to a standardized format (H W = 320 320 pixels), random horizontal flipping with probability p = 0.5, random rotations in the range of 15 based on an uniform probability distribution and normalization. For validation and testing, only resizing and normalization were applied.