Differentiable Knowledge Units Improve Image Recognition

A novel technique has been introduced for incorporating domain expertise into deep neural networks, aimed at enhancing image recognition accuracy. This method features a Differentiable Knowledge Unit (DKU) that adjusts classifier logits to yield improved class probabilities. In contrast to traditional approaches that embed prior knowledge within loss functions or in post-processing, the DKU autonomously extracts symbolic knowledge from main task supervision, eliminating the need for concept labels. It employs implication rules to illustrate relationships between task classes and implicit concepts derived solely from the data. Special classifiers determine concepts, and their probabilities are integrated with primary class probabilities into the DKU. Through fuzzy inference, the DKU generates a logic-based adjustment vector to modify primary class logits, effectively tackling the issue of missing symbolic rules in practical vision applications. This study is available on arXiv with the identifier 2604.27759.