The paper presents a dynamic few-shot visual learning system that can efficiently learn novel categories from a few training examples while maintaining accuracy on the base categories. The system is designed to mimic the human visual system's ability to learn new concepts from limited examples. Key contributions include: 1. **Attention-based Few-Shot Classification Weight Generator**: This component generates classification weight vectors for novel categories using a few training examples, leveraging past knowledge about the visual world through an attention mechanism. 2. **Cosine Similarity-based ConvNet Classifier**: This classifier unifies the recognition of both base and novel categories by computing cosine similarities between feature representations and classification weight vectors, leading to better generalization on novel categories. The system is evaluated on the Mini-ImageNet dataset and a recent few-shot benchmark, achieving state-of-the-art results in both few-shot recognition and the ability to not forget base categories. The code and models are available at <https://github.com/gidariss/FewShotWithoutForgetting>.The paper presents a dynamic few-shot visual learning system that can efficiently learn novel categories from a few training examples while maintaining accuracy on the base categories. The system is designed to mimic the human visual system's ability to learn new concepts from limited examples. Key contributions include: 1. **Attention-based Few-Shot Classification Weight Generator**: This component generates classification weight vectors for novel categories using a few training examples, leveraging past knowledge about the visual world through an attention mechanism. 2. **Cosine Similarity-based ConvNet Classifier**: This classifier unifies the recognition of both base and novel categories by computing cosine similarities between feature representations and classification weight vectors, leading to better generalization on novel categories. The system is evaluated on the Mini-ImageNet dataset and a recent few-shot benchmark, achieving state-of-the-art results in both few-shot recognition and the ability to not forget base categories. The code and models are available at <https://github.com/gidariss/FewShotWithoutForgetting>.