Hypernetwork based Continual Zero-Shot Learning for Efficient Knowledge Transfer

Continual learning (CL) and zero-shot learning (ZSL) are two major paradigms of machine learning that solve complex problems. CL targets at training models to learn from continuous data streams without forgetting the knowledge acquired from earlier tasks while ZSL targets at training models to acquire knowledge to classify both seen and unseen classes during testing. Continual zero shot learning (CZSL) combines both CL and ZSL such that model should be capable to learn over sequence of tasks but also able to transfer knowledge to classify unseen classes. Despite various existing works able to solve CZSL, many approaches often fail to transfer knowledge across the tasks while also maintaining parameter efficiency. To address these limitations, we aim to propose novel hypernetwork based continual zero shot learning (HCZSL) that able to transfer knowledge across tasks with reduced parameters. Hypernetwork generate weights of main network i.e., Variational Auto-Encoder (VAE) with the help of learnable task embeddings. This enables model to transfer knowledge to new tasks while also dealing with forgetting of previously acquired knowledge. These task embeddings are very small learnable parameters that are capable to store task-specific information. Through experiments, we demonstrate the effectiveness of our proposed HCZSL in its ability to deal with forgetting while also maintaining parameter efficiency.