[Paper Review] Small Sample Learning in Big Data Era

Jun Shu, Zongben Xu+1 more|arXiv (Cornell University)|Aug 14, 2018

Machine Learning and Algorithms274 references56 citations

TL;DR

A survey of Small Sample Learning (SSL) techniques, distinguishing experience learning and concept learning, and outlining related methods, neuroscience foundations, challenges, and future directions.

ABSTRACT

As a promising area in artificial intelligence, a new learning paradigm, called Small Sample Learning (SSL), has been attracting prominent research attention in the recent years. In this paper, we aim to present a survey to comprehensively introduce the current techniques proposed on this topic. Specifically, current SSL techniques can be mainly divided into two categories. The first category of SSL approaches can be called "concept learning", which emphasizes learning new concepts from only few related observations. The purpose is mainly to simulate human learning behaviors like recognition, generation, imagination, synthesis and analysis. The second category is called "experience learning", which usually co-exists with the large sample learning manner of conventional machine learning. This category mainly focuses on learning with insufficient samples, and can also be called small data learning in some literatures. More extensive surveys on both categories of SSL techniques are introduced and some neuroscience evidences are provided to clarify the rationality of the entire SSL regime, and the relationship with human learning process. Some discussions on the main challenges and possible future research directions along this line are also presented.

Motivation & Objective

• Define SSL and motivate its importance in the big data era.
• Distinguish and describe the two SSL branches: concept learning and experience learning.
• Summarize representative techniques for concept learning and their connections to prior work.
• Summarize representative techniques for experience learning and how augmented data and knowledge can compensate small samples.
• Discuss challenges, neuroscience evidence, and future research directions in SSL.

Proposed method

• Present a formal definition of SSL and its two learning categories.
• Describe a general methodology for concept learning including intension/extension matching and new concept formation.
• Survey intension matching approaches from visual-to-semantic mappings and semantic relatedness.
• Outline the role of augmented data and knowledge systems in experience learning.
• Relate SSL to cognitive science concepts and provide neuroscience evidence.
• Discuss long-tail, data scarcity, and weak/webly supervision as SSL motivation.

Experimental results

Research questions

• RQ1 What are the core definitions and classifications within Small Sample Learning (SSL)?
• RQ2 What techniques constitute concept learning and experience learning in SSL?
• RQ3 How can SSL leverage representations, mappings, and knowledge to operate with few samples?
• RQ4 What evidence from neuroscience supports SSL and how does it relate to human learning?
• RQ5 What are the main challenges and future directions for SSL in the big data era.

Key findings

• SSL is divisible into concept learning and experience learning, enabling recognition, generation, and reasoning from few samples.
• Experience learning uses augmented data and knowledge systems to compensate for limited data, while concept learning relies on matching rules between concepts and small samples.
• Intension matching maps between visual/semantic representations to align concepts with data, enabling zero-shot and few-shot tasks.
• Semantic embedding and semantic relatedness approaches enable transferring knowledge from seen to unseen classes in zero-shot/few-shot settings.
• Neuroscience concepts such as episodic memory, imagination, and compositionality provide a rationale for rapid learning with prior knowledge.
• The paper discusses challenges like weak supervision, long-tail distributions, and data scarcity, proposing SSL as a path toward more human-like learning.