This paper proposes a dynamic few-shot visual learning system that can learn novel categories from only a few examples while not forgetting the base categories it was trained on. The system is designed to mimic the human visual system's ability to learn new concepts quickly and recognize them accurately. The key contributions include: (1) a few-shot object recognition system that dynamically learns novel categories without forgetting the base ones; (2) an attention-based few-shot classification weight generator that generates classification weight vectors for novel categories based on past knowledge; and (3) a cosine-similarity based ConvNet classifier that unifies the recognition of both base and novel categories and improves generalization on unseen categories. The system is evaluated on the Mini-ImageNet dataset and the recently introduced few-shot benchmark by Bharath and Girshick, achieving state-of-the-art results. The approach outperforms prior methods in both few-shot recognition performance and base category recognition accuracy. The system is implemented using a ConvNet-based recognition model with a cosine-similarity based classifier and a few-shot classification weight generator that dynamically generates classification weight vectors for novel categories. The framework is trained on a set of base categories with a large amount of training data, and during test time, it uses the few-shot classification weight generator to learn novel categories from only a few examples. The system is able to recognize both base and novel categories in a unified manner without sacrificing the accuracy of the base categories. The results show that the proposed approach significantly improves the few-shot recognition performance on both base and novel categories compared to prior methods.This paper proposes a dynamic few-shot visual learning system that can learn novel categories from only a few examples while not forgetting the base categories it was trained on. The system is designed to mimic the human visual system's ability to learn new concepts quickly and recognize them accurately. The key contributions include: (1) a few-shot object recognition system that dynamically learns novel categories without forgetting the base ones; (2) an attention-based few-shot classification weight generator that generates classification weight vectors for novel categories based on past knowledge; and (3) a cosine-similarity based ConvNet classifier that unifies the recognition of both base and novel categories and improves generalization on unseen categories. The system is evaluated on the Mini-ImageNet dataset and the recently introduced few-shot benchmark by Bharath and Girshick, achieving state-of-the-art results. The approach outperforms prior methods in both few-shot recognition performance and base category recognition accuracy. The system is implemented using a ConvNet-based recognition model with a cosine-similarity based classifier and a few-shot classification weight generator that dynamically generates classification weight vectors for novel categories. The framework is trained on a set of base categories with a large amount of training data, and during test time, it uses the few-shot classification weight generator to learn novel categories from only a few examples. The system is able to recognize both base and novel categories in a unified manner without sacrificing the accuracy of the base categories. The results show that the proposed approach significantly improves the few-shot recognition performance on both base and novel categories compared to prior methods.