This paper considers incremental few-shot learning, which requires a model to continually recognize new categories with only a few examples provided. Our study shows that existing methods severely suffer from catastrophic forgetting, awell-known problem in incremental learning, which is aggravated due to data scarcity andimbalance inthefew-shot setting.

Firstly, the adaptiveknowledge integration mechanism isdesigned tolearn better representations from all heterogeneous models with deviated representation abilities.

These forachievingT requires onthepriorB> 1. 2.1 Generalizing Choosing Follow [Kalaiand worst-case FPLalgorithm Follow (FPML).

Circles are nodes, and rectangles are VQ codewords. A double circle indicates nodes in the current mini-batch.

End-to-end learnable discrete-continuous models are compositional, tend togeneralize better,and are more interpretable.