
Intrinsically interpretable (deep learning) methods aim to bridge the gap between accuracy and interpretability. Their idea is to combine deep representation learning with easily understandable decision layers to construct a model with a traceable reasoning process. Prominent examples are ProtoPNet, Concept Bottleneck Models, B-Cos, and their derivatives. These decision models are designed to transparently reveal the logic behind their predictions during inference. Despite their advantages, intrinsically interpretable models require unique architectures and training procedures, and they can show a drop in performance compared to their black-box counterparts. Furthermore, learning prototypes that mimic human reasoning is still an open challenge. This track aims to explore the latest challenges and developments in intrinsically interpretable models, including evaluation techniques, the interpretability-accuracy trade-off, theoretical foundations, their practical applications, and their broader impact on society.
- Intrinsically interpretable architectures (eg. ProtoPNet, PIPNet, B-cos)
- Intrinsically interpretable representation learning (eg. ProtoMetric)
- Intrinsically interpretable generative models (eg. ProtoFlow)
- Evaluation of intrinsically interpretable models (eg. Spatial Misalignment Benchmark)
- XAI evaluation benchmarks, frameworks, or contributions to existing benchmarks (e.g. FunnyBirds)
- Applications of intrinsically interpretable models (eg. XProtoNet, Proto-BagNets)
- Intrinsically interpretable models for different data modalities (eg. ProGReST, “This Reads Like That”)
- User-centred aspects of intrinsically interpretable models (eg. “Help Me Help the AI”)
- Theory of intrinsically interpretable models
- Human-adaptable intrinsically interpretable models (eg. R3-ProtoPNet)
- Interactive intrinsically interpretable models (eg. ProtoPDebug)
- Relation of intrinsic interpretability and mechanistic interpretability
- Intrinsic interpretability for AI safety
- Intrinsic interpretability in the context of current AI challenges (eg. ICICLE for Continual Learning)







