IDEA #0IR8HM Portal Learning: An approach to out-of-distribution problem in machine learning. 5/14/2021 17:22 DA.

21A0022; DA, 5/14/2021 17:22 Software NTD Tian (Beatrice) Cai, Graduate Center, Computer Science Lei Xie, Hunter College, Computer Science (primary contact) One of fundamental challenges in applying machine learning (ML) to real-world problems is out-of-distribution (OOD) problems, i.e., the distribution of unseen data for the task of interest (unlabeled data) is significantly different from that of the data used in the training of ML models (labeled data). Although causal learning is a theoretically sound solution, it is still infeasible to solve the real-world problem where data is huge, high-dimensional, noisy, heterogenous, and biased. We invent a novel approach, portal learning, to address real-world OOD problems. Portal learning consists of three conceptual steps: (1) obtain a global landscape of feature space by using both labeled and unlabeled data. (2) connect labeled data to unlabeled data by identifying similar regions between unlabeled data and labeled data. (3) explore unique regions of unlabeled data starting from multiple origins of labeled data through the similar regions. Portal learning unifies several latest advances in deep learning, particularly, self-supervised learning, transfer learning, multi-task learning, meta-learning, and semi-supervised learning. It establishes a new paradigm for machine learning.