Design a Flexible Multi-Class Classification Model

Design Problem: Build a model/architecture to classify 10 classes today, where classes may be added or removed in the future - the model architecture itself should not need to change. Requirements: Support a variable/open number of output classes without retraining or redesigning the network from scratch. Design: Favor an embedding-based/metric-learning approach (e.g. learn a shared embedding space and classify via nearest-class-prototype or similarity to class embeddings) or a two-stage feature-extractor + swappable classifier-head design, so new classes can be added by adding prototypes/fine-tuning only the head rather than the whole network. Discuss trade-offs vs a fixed-size softmax output layer (which requires re-architecting and full retraining whenever the class count changes).