Understanding Inter-Concept Relationships in Concept-Based Models

Authors: Naveen Janaki Raman, Mateo Espinosa Zarlenga, Mateja Jamnik

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental First, we empirically demonstrate that state-of-the-art concept-based models produce representations that lack stability and robustness, and such methods fail to capture inter-concept relationships. Then, we develop a novel algorithm which leverages inter-concept relationships to improve concept intervention accuracy, demonstrating how correctly capturing inter-concept relationships can improve downstream tasks.
Researcher Affiliation Academia Naveen Raman 1 Mateo Espinosa Zarlenga 2 Mateja Jamnik 2 1Carnegie Mellon University 2University of Cambridge.
Pseudocode Yes Algorithm 1 Basis Aided Concept Intervention
Open Source Code Yes Our code is available here: https://github.com/ naveenr414/Concept-Learning.
Open Datasets Yes Coloured MNIST (Arjovsky et al., 2019), d Sprites (Matthey et al., 2017), CUB (Wah et al., 2011), Che Xpert (Irvin et al., 2019)
Dataset Splits Yes For the d Sprites and Che Xpert datasets, we use 2,500 data points for the training, and 750 for validation and testing. For the MNIST dataset, we use 60,000 data points for training and 10,000 data points for validation. For CUB, we use 4,796 data points for training, 1,198 for validation, and 5,794 data points for testing.
Hardware Specification Yes We run our GPU experiments on either an NVIDIA TITAN Xp with 12 GB of GPU RAM on Ubuntu 20.04, or NVIDIA A100-SXM, using at most 8 GB of GPU with Red Hat Linux 8.
Software Dependencies No For concept intervention experiments we use the Py Torch library (Paszke et al., 2019). (Does not include version number)
Experiment Setup Yes We train models for 25, 50, and 100 epochs and measure the impact of label bases on concept intervention accuracy.