Concept probing has recently garnered increasing interest as a way to help interpret artificial neural networks, dealing both with their typically large size and their subsymbolic nature, which ultimately renders them unfeasible for direct human interpretation. Concept probing works by training additional classifiers to map the internal representations of a model into human-defined concepts of interest, thus allowing humans to peek inside artificial neural networks. Research on concept probing has mainly focused on the model being probed or the probing model itself, paying limited attention to the data required to train such probing models. In this paper, we address this gap. Focusing on concept probing in the context of image classification tasks, we investigate the effect of the data used to train probing models on their performance. We also make available concept labels for two widely used datasets.

Concept probing has recently gained popularity as a way for humans to peek into what is encoded within artificial neural networks. In concept probing, additional classifiers are trained to map the internal representations of a model into human-defined concepts of interest. However, the performance of these probes is highly dependent on the internal representations they probe from, making identifying the appropriate layer to probe an essential task. In this paper, we propose a method to automatically identify which layer’s representations in a neural network model should be considered when probing for a given human-defined concept of interest, based on how informative and regular the representations are with respect to the concept. We validate our findings through an exhaustive empirical analysis over different neural network models and datasets.

Neural networks have been the key to solve a variety of different problems. However, neural network models are still regarded as black boxes, since they do not provide any human-interpretable evidence as to why they output a certain result. We address this issue by leveraging on ontologies and building small classifiers that map a neural network model's internal state to concepts from an ontology, enabling the generation of symbolic justifications for the output of neural network models. Using an image classification problem as testing ground, we discuss how to map the internal state of a neural network to the concepts of an ontology, examine whether the results obtained by the established mappings match our understanding of the mapped concepts, and analyze the justifications obtained through this method.