Attributing Learned Concepts in Neural Networks to Training Data
This addresses the problem of understanding concept formation for trustworthy AI, but it is incremental as it builds on existing attribution and probing techniques.
The study investigated how specific training examples influence the learning of human-interpretable concepts in neural networks by combining data attribution and concept probing methods, finding that removing top-attributing images did not alter concept location or sparsity, suggesting diffuse feature reliance across examples.
By now there is substantial evidence that deep learning models learn certain human-interpretable features as part of their internal representations of data. As having the right (or wrong) concepts is critical to trustworthy machine learning systems, it is natural to ask which inputs from the model's original training set were most important for learning a concept at a given layer. To answer this, we combine data attribution methods with methods for probing the concepts learned by a model. Training network and probe ensembles for two concept datasets on a range of network layers, we use the recently developed TRAK method for large-scale data attribution. We find some evidence for convergence, where removing the 10,000 top attributing images for a concept and retraining the model does not change the location of the concept in the network nor the probing sparsity of the concept. This suggests that rather than being highly dependent on a few specific examples, the features that inform the development of a concept are spread in a more diffuse manner across its exemplars, implying robustness in concept formation.