Jiaqi Wang

Most works of interpretable neural networks strive for learning the semantics concepts merely from single modal information such as images. However, humans usually learn semantic concepts from multiple modalities and the semantics is encoded by the brain from fused multi-modal information. Inspired by cognitive science and vision-language learning, we propose a Prototype-Concept Alignment Network (ProCoNet) for learning visual prototypes under the guidance of textual concepts. In the ProCoNet, we have designed a visual encoder to decompose the input image into regional features of prototypes, while also developing a prompt generation strategy that incorporates in-context learning to prompt large language models to generate textual concepts. To align visual prototypes with textual concepts, we leverage the multimodal space provided by the pre-trained CLIP as a bridge. Specifically, the regional features from the vision space and the cropped regions of prototypes encoded by CLIP reside on different but semantically highly correlated manifolds, i.e. follow a multi-manifold distribution. We transform the multi-manifold distribution alignment problem into optimizing the projection matrix by Cayley transform on the Stiefel manifold. Through the learned projection matrix, visual prototypes can be projected into the multimodal space to align with semantically similar textual concept features encoded by CLIP. We conducted two case studies on the CUB-200-2011 and Oxford Flower dataset. Our experiments show that the ProCoNet provides higher accuracy and better interpretability compared to the single-modality interpretable model. Furthermore, ProCoNet offers a level of interpretability not previously available in other interpretable methods.

When humans explain their reasoning, such as their classification decisions, they often break down an image into parts and highlight the evidence from those parts to support the concepts they have in mind. Drawing inspiration from this cognitive process, several self-explaining models have been proposed to explain predictions by part-level concepts. However, these models can be limited by their structure and difficulty in determining the effect of individual parts on the output category. To address these challenges, we introduce a self-explaining architecture that uses a plug-in transparent embedding space (TesNet) to connect high-level input patches (e.g. feature maps or tokens) with output categories. The transparent embedding space is spanned by basis concepts and constructed on the Grassmann manifold. The basis concepts are enforced to be category-aware, and within-category concepts are orthogonal to each other, ensuring the embedding space is disentangled. To reduce concept redundancy and restore the concept space structure, we introduce two concept pruning methods and a new re-training strategy to build a slimming transparent embedding space. We verify the scalability of TesNet through experiments on deep networks such as VGG, ResNet, DenseNet, and Vision Transformer. Additionally, we design several metrics for self-explaining models to quantify interpretability and compare them with state-of-the-art self-explaining methods. Our experiments demonstrate that TesNet is much more effective for classification tasks, providing better interpretability on predictions and improving final accuracy.