Few-shot Classification with First-order Task Agnostic Meta-learning

Meta-learning approaches are typically used to solve few-shot learning tasks by training on a variety of data in the hopes of developing generalization abilities that can be applied to new tasks. Model-Agnostic Meta-Learning (MAML) is one of the popular ways for few-shot learning currently. However, the generalizability is impacted when the meta-learner is over-trained in the meta-training, leading to a bias toward existing tasks. Besides, MAML involves a second-order gradient, which costs a lot. In this paper, we propose a First-order Task-agnostic Meta-learning algorithm (TA-Reptile). Our entropy-based method utilizes a first-order gradient update to learn an unbiased model with a fine set of initialization parameters, aiming to address the difficulty of overperforming in classification tasks. Experiments on benchmark datasets illustrate that TA-Reptile has competitive performance in few-shot classification tasks.