Few-Shot Lifelong Learning

Many real-world classification problems often have classes with very few labeled training samples. Moreover, all possible may not be initially available for training, and given incrementally. Deep learning models need to deal this two-fold problem in order perform well real-life situations. In paper, we propose a novel Few-Shot Lifelong Learning (FSLL) method that enables deep lifelong/continual on few-shot data. Our selects parameters from the model every new set of instead full model. This helps preventing overfitting. We choose such way only currently unimportant get selected. By keeping important intact, our approach minimizes catastrophic forgetting. Furthermore, minimize cosine similarity between old class prototypes maximize their separation, thereby improving performance. also show integrating self-supervision improves performance significantly. experimentally significantly outperforms existing methods miniImageNet, CIFAR-100, CUB-200 datasets. Specifically, outperform state-of-the-art by an absolute margin 19.27% CUB dataset.