Avoiding Catastrophic Forgetting in Visual Classification Using Human Concept Formation

This work networks can exceed human capabilities in certain tasks aims to combine computer vision principles with this prior such as object detection and classification (He, Zhang, Ren, & approach and explore the idea of incorporating new visual information Sun, 2016). However, such networks cannot handle continual incrementally without erasing previously learned learning of new tasks without forgetting previously learned data. Our results demonstrate that Cobweb/4V does not exhibit data. Catastrophic forgetting is a fundamental challenge for catastrophic forgetting, only limited interference effects artificial neural networks (McCloskey & Cohen, 1989). This when compared to neural networks. We find that Cobweb/4V phenomenon happens when the network is trained on multiple is competitive with neural network approaches while having tasks sequentially, and to meet the objective of the new minimal forgetting effects. It is also more data efficient task, it changes the weights learned to perform the previous and achieves asymptotic performance with fewer examples.