- Incremental Residual Idea Bottleneck Fashions(arXiv)
Authors: Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, Yuwang Wang
Summary: Idea Bottleneck Fashions (CBMs) map the black-box visible representations extracted by deep neural networks onto a set of interpretable ideas and use the ideas to make predictions, enhancing the transparency of the decision-making course of. Multimodal pre-trained fashions can match visible representations with textual idea embeddings, permitting for acquiring the interpretable idea bottleneck with out the experience idea annotations. Current analysis has targeted on the idea financial institution institution and the high-quality idea choice. Nonetheless, it’s difficult to assemble a complete idea financial institution by people or massive language fashions, which severely limits the efficiency of CBMs. On this work, we suggest the Incremental Residual Idea Bottleneck Mannequin (Res-CBM) to handle the problem of idea completeness. Particularly, the residual idea bottleneck mannequin employs a set of optimizable vectors to finish lacking ideas, then the incremental idea discovery module converts the complemented vectors with unclear meanings into potential ideas within the candidate idea financial institution. Our method will be utilized to any user-defined idea financial institution, as a post-hoc processing methodology to reinforce the efficiency of any CBMs. Moreover, to measure the descriptive effectivity of CBMs, the Idea Utilization Effectivity (CUE) metric is proposed. Experiments present that the Res-CBM outperforms the present state-of-the-art strategies when it comes to each accuracy and effectivity and achieves comparable efficiency to black-box fashions throughout a number of datasets.
