Multi-Instance Learning with Incremental Classes

In recent years, multi-instance learning (MIL) has been widely used in complicated data problems, but the existing MIL methods often study a fixed number of categories in a closed environment. However, in real applications, novel categories are constantly added to the system, such as the continuous emergence of new topics in the development of science or social media. Due to storage restrictions or confidentiality agreements, old data may become invisible over time, which makes the model forget the previously learned knowledge when directly learning new categories. Incremental learning is often used to deal with the aforementioned problems. The mining of multi-instance learning with incremental classes is very meaningful, but the current works on this is rare to be focused. We propose a novel multi-instance incremental data mining method based on both attention mechanism and prototype classifier mapping. Through the attention mechanism, the MIL bags are selectively merged into unified feature representations, which will be used to generate the corresponding storable category prototypes. Through the prototype classifier mapping, each category prototype is mapped into an unbiased and robust classifier. The prediction results of the classifier generated in the previous incremental stage are used to perform knowledge distillation on the prediction results of the classifier generated in novel incremental stages, so that the model can retain the old knowledge with very low storage under MIL. Experimental results on benchmarks of three different tasks show that our proposed method have achieved effective performance in MIL with incremental classes.