Mathilde Bonin and Sergio Mover
École Polytechnique, Palaiseau, France
Neural networks for vision are widely used but do not provide explanations for their decisions. A recent approach (i.e., ConSpec [1]) uses concepts to bridge the gap between pixel-like features and human-understandable text, and a propositional logic to express specifications of the model’s behavior. However, such approach remains limited and cannot express simple properties involving multiple concepts, predicates on their strength (a proxy for the presence of a concept in the input), or predicates on multiple input images.
This limitation is mostly due to the non-linear relationship between concepts’ strengths and the latent space representation of the model, which makes satisfiability checking difficult and restricts the application of the logic to explain the network’s decisions. We propose Concept Part Network (ConPNet), a novel architecture that integrates the concepts’ strengths computation directly into a classifier and then uses these strengths for prediction. Such architecture allows us to extend ConSpec to support arbitrary linear combinations of concepts’ strengths (ConSpecReal), which are necessary to express a wider range of specifications, and efficiently check them with an SMT solver.
Our initial experiments show that ConPNet retains competitive classification accuracy and supports effective explanation of predictions through ConSpecReal formulas.
1. Ravi Mangal et al.: Concept-based analysis of neural networks via vision-language models. In: International Symposium on AI Verification. Springer. 2024, pp. 49–77.