[Paper Review] Learning from Few Samples: A Survey

Nihar Bendre, Hugo Terashima‐Marín+1 more|arXiv (Cornell University)|Jul 30, 2020

Domain Adaptation and Few-Shot Learning140 references37 citations

TL;DR

This survey analyzes few-shot meta-learning techniques in computer vision, proposing a taxonomy and comparing methods on Omniglot and Mini-Imagenet benchmarks.

ABSTRACT

Deep neural networks have been able to outperform humans in some cases like image recognition and image classification. However, with the emergence of various novel categories, the ability to continuously widen the learning capability of such networks from limited samples, still remains a challenge. Techniques like Meta-Learning and/or few-shot learning showed promising results, where they can learn or generalize to a novel category/task based on prior knowledge. In this paper, we perform a study of the existing few-shot meta-learning techniques in the computer vision domain based on their method and evaluation metrics. We provide a taxonomy for the techniques and categorize them as data-augmentation, embedding, optimization and semantics based learning for few-shot, one-shot and zero-shot settings. We then describe the seminal work done in each category and discuss their approach towards solving the predicament of learning from few samples. Lastly we provide a comparison of these techniques on the commonly used benchmark datasets: Omniglot, and MiniImagenet, along with a discussion towards the future direction of improving the performance of these techniques towards the final goal of outperforming humans.

Motivation & Objective

• Motivate the study of learning from few samples and highlight the challenge of limited labeled data for novel categories.
• Provide a taxonomy of few-shot meta-learning approaches across four categories.
• Summarize seminal works and their strategies within each category.
• Compare performance of techniques on standard CV benchmarks and discuss future directions.

Proposed method

• Classify few-shot meta-learning techniques into data-augmentation, embedding, optimization, and semantic-based groups.
• Describe representative seminal works in each category and their core ideas.
• Explain how these techniques address learning from a few labeled samples for novel tasks.
• Compare methods using common benchmark datasets Omniglot and Mini-Imagenet.
• Discuss connections to transfer learning and self-supervised learning as complementary approaches.

Experimental results

Research questions

• RQ1 What are the main categories of few-shot meta-learning techniques in computer vision and their core ideas?
• RQ2 How do data-augmentation, embedding, optimization, and semantic-based methods perform on standard benchmarks like Omniglot and Mini-Imagenet?
• RQ3 What are the limitations of current approaches and potential future directions for improving few-shot learning performance?

Key findings

• The paper provides a taxonomy of four categories: data-augmentation, embedding, optimization, and semantic-based learning for few-shot, one-shot, and zero-shot settings.
• It reviews seminal works within each category and their approaches to learning from limited samples.
• It compares techniques on Omniglot and Mini-Imagenet datasets to discuss relative strengths and limitations.
• The survey discusses future directions toward surpassing human performance on few-shot learning tasks.