SoFTNet: A concept-controlled deep learning architecture for interpretable image classification

Interpreting deep learning (DL)-based computer vision models is challenging due to the complexity of internal representations. Most recent techniques for rendering DL learning outcomes interpretable operate on low-level features rather than high-level concepts. Methods that explicitly incorporate high-level concepts do so through a determination of the relevancy of user-defined concepts or else concepts extracted directly from the data. However, they do not leverage the potential of concepts to explain model predictions. To overcome this challenge, we introduce a novel DL architecture – the Slow/Fast Thinking Network (SoFTNet) – enabling users to define/control high-level features and utilize them to perform image classification predicatively. We draw inspiration from the dual-process theory of human thought processes, decoupling low-level, fast & non-transparent processing from high-level, slow & transparent processing. SoFTNet hence uses a shallow convolutional neural network for low-level processing in conjunction with a memory network for high-level concept-based reasoning. We conduct experiments on the CUB-200-2011 and STL-10 datasets and also present a novel concept-based deep K-nearest neighbor approach for baseline comparisons. Our experiments show that SoFTNet achieves comparable performance to state-of-art non-interpretable models and outperforms comparable interpretative methods.