Adaptive lifelong learning

Date

2018-12-20

Authors

Parul

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Lifelong learning is an emerging field in machine learning that still requires a lot of research. In lifelong learning, the tasks are presented sequentially, the system learns knowledge at each task and the goal is to retain the learned knowledge and utilize it when learning a new task. Exponentially Weighted Aggregation for Lifelong Learning (EWA-LL) is a meta-algorithm used in lifelong learning setting. It transfers information from previous tasks to the next. A prior distribution is maintained on the set of representations, which is updated after the encounter of each new task using the exponentially weighted aggregation (EWA) procedure. This project tries to relax the problem and explores the case of an easy scenario where we have some more information about the data. It implements adaptive learning in lifelong learning setting. It utilizes the adaptive learning algorithm Follow The Leader with Dropout Perturbations (FTL-DP) used in Online Prediction with Expert Advice. FTL-DP sets the losses of the experts to 0 or 1 at each task based on the dropout probability before selecting the leader. This project transports FTL-DP to lifelong learning setting. The goal is to prove that adaptive algorithm in lifelong learning is a better approach than EWA-LL as it gives smaller regret for certain easy problems while still maintaining the regret bounds similar to EWA-LL for the harder problems.

Description

Keywords

Lifelong learning, Machine Learning, Adaptive learning

Citation