Meta-Learning 4th Ed

About the Course: We will cover topics and recent research papers focusing on the broad goal of developing autonomous intelligent agents that generalise beyond the data they were trained on, are able to rapidly adapt to new situations with limited experience, and can learn continuously. Topics covered will include the basic metric, optimisation and model based meta-learning techniques for making predictions as well as decisions. We will also cover recent generative models, such as large language models and diffusion models, examining their potential to aid in rapidly adapting to novel tasks.

Note: Advanced deep-learning concepts needed for the above, such as recurrent networks, reinforcement learning and transformers will also be introduced along the way, in a hands-on manner. However, prior hands-on exposure to gradient-based deep-learning using pytorch is an essential pre-requisite.

We will study a few seminal/recent papers spanning the above broad swath of topics in depth, implementing the key ideas involved via a in-class of programming exercises, programming assignments, and a course project involving implementing a research paper idea within an autonomous intelligent agent architecture . Data and base code for the assignments and project will be provided, drawn from visual reasoning and decision-making tasks. Each course project is expected to involve/lead-to research explorations that may result in opportunities for publication in the future.

Topics covered in lectures (live but hybrid) will be as follows:

1. Deep-learning Refresher: MLPs

2. Autonomous Intelligent Agents

3. Few-shot Learning (FSL) and Metric/Optimization-based Meta-learning

4. Introduction to (Deep) Reinforcement Learning and Meta Reinforcement Learning

5. Model-based Meta-learning, Online and Continual Learning

6. Introduction to Transformers and Decision Transformers

7. Introduction to (Large) Language Models(LLMs) and FSL using LLMs

8. Neuro-evolutionary Learning

9. Diffusion-based Generative Models and Meta-learning Opportunities using Diffusion Models

Pre-requisite: Familiarity with object-oriented programming in python and exposure to machine-learning and deep-learning with pytorch are mandatory pre-requisites, and will be required for the assignments/projects. Familarity with pandas is also desirable but not mandatory.