[2J5-GS-2] Machine learning: Advancement reinforcement learning (1)

[2J5-GS-2-02] Continual Learning Based on Amortized Inference

〇Hirono Kawashima¹, Makoto Kawano², Wataru Kumagai³, Kota Matsui³, Jin Nakazawa¹ (1. Keio University , 2. The University of Tokyo , 3. RIKEN AIP)

Keywords:Continual Learning, Neural Process

In continual learning, studies are being actively conducted on techniques that can cope with an increase in tasks while preventing catastrophic forgetting, in which the accuracy of past tasks drops significantly when learning multiple tasks sequentially.
In this study, we propose an Continual Amortized Learning Model (CALM) based on the structure of the Neural Process as a new method that saves the network of past tasks and does not learn by adding training data.
CALM consists of two neural networks: Task Weight Encoder, that calculates the task-specific weight, and Feature Extractor, that extracts the features of input data.
By applying task-specific weights to the features of the input image, task-specific outputs are possible while using a common network for all the tasks.
In the experiment, we worked on task incremental learning of Split-MNIST, and verified that the task accuracy was maintained even when learning was performed sequentially using the proposed method.