Deep learning models are dominating pattern recognition, machine learning and computer vision, and have achieved human-level performances in various tasks, such as image classification. The success and unprecedented performances of state-of-the-art learning models are often achieved via training on large-scale labeled data sets. Nevertheless, current models still have difficulty generalizing to novel tasks (e.g., new image classes) unseen during training, given only a few labeled instances for these new tasks. In contrast, humans can learn new tasks easily from a few examples, by leveraging prior experience and related context. Few-shot learning has emerged as an appealing paradigm to bridge this gap.

Few-shot learning has recently triggered substantial research efforts and interests, with large numbers of publications within the machine learning, computer vision and pattern recognition communities. For instance, meta-learning or “learning-to-learn” approaches, which create artificial few-shot tasks during training to mimic generalization difficulties, have attracted widespread attention within the community.

This tutorial discusses the recent progress made in few-shot learning, juxtaposing meta-learning and advanced transfer-learning approaches. It will also discuss various applications of few-shot learning in pattern recognition and computer vision. We will emphasize recent trends, and insights from the literature, and point to important limitations in current few-shot models and benchmarks. We will focus on the basics. Understanding of few-shot concepts and models so that tutorial attendees, without prior experience in few-shot learning, could extend/use effectively few-shot methods and publicly available code.

The tutorial will be divided into 4 parts + conclusion, corresponding to the following tentative plan:

1. Introduction to few-shot learning

   1. Motivation

   2. Basic few-shot settings and definitions

   3. Commonly tackled tasks

   4. Standard benchmarks and their limitations

2. Meta-learning approaches

   1. Metric-learning based approaches (e.g, ProtoNets)

   2. Optimization-based approaches (e.g., MAML)

3. Advanced transfer-learning approaches

   1. Inductive vs. transductive inference

   2. Strong baselines (e.g., SimpleShot)

   3. Manifold-regularization approaches (e.g., LaplacianShot)

   4. Information-theoretic approaches (e.g., TIM)

4. An example of advanced application: Few-shot segmentation

   1. Specific challenges

   2. Meta-learning vs. transfer-learning approaches: A case study of transductive few-shot inference

   3. Limitations of current models and benchmarks

5. Conclusion and outlook

This tutorial has the objective to be as inclusive as possible in its targeted audience. As such, we divide the learning objective into two groups, the "basic objectives" targeted at everyone, and the "advanced objectives" targeted at a more expert audience (graduate students, researchers, practitioners).

Basic objectives

• • • Understand the motivation behind the few-shot learning problem

    • Understand the few-shot setting, and be able to contextualize it in the Machine Learning spectrum

    • Get a sense of the breadth of applications for Few-Shot Learning

    • Have an overview of the litterature, and understand the popular approaches (meta-learning, transfer learning)

Advanced objectives

• • • Understand the practical limitations of meta-learning approaches

    • Understand the difference between induction and transduction, and their practical instantiation in the context of Few-Shot learning

    • Understand the recent advances in model-agnostic transductive learning

    • Have an overview of important emerging research directions for Few-Shot Learning