Amortized Bayesian Meta-Learning with Accelerated Gradient Descent Steps

Recent meta-learning models often learn priors from observed tasks using a network optimized via stochastic gradient descent (SGD), which usually takes more training steps to convergence. In this paper, we propose an accelerated Bayesian structure with inference (ABML-SIN). The proposed model aims solve the procedure of improve speed and efficiency. Current approaches hardly converge within few steps, owing small number samples. Therefore, introduce learning based on teacher–student architecture meta-latent variable θt for task t. With amortized fast network, meta-learner is able task-specific latent steps; thus, it improves meta-learner. To refine variables generated transductive amortization meta-learner, SIN—followed by conventional SGD-optimized network—is introduced as student–teacher online-update parameters. SIN extracts local accelerates convergence network. Our experiments simulation data demonstrate that method provides generalization scalability unseen samples, produces competitive/superior uncertainty estimations few-shot two widely adopted 2D datasets fewer epochs compared state-of-the-art approaches. Furthermore, parameters act perturbations weights, enhancing probability accelerating efficiency Extensive qualitative show our performs well across different in both simulated real-world circumstances.