Meta-Learning 3rd Ed

IIT Delhi CSE: COL870 Special Topics in Machine Learning

IIIT Delhi CSE: CSE663A Meta-Learning

BITS Goa CS: BITS G513 (Study in Advanced Topics)

Meta-Learning 3rd Ed

3 credits @BITS-Goa, IIITD and IITD

Spring 2022

Instructor:

Dr. Gautam Shroff

(TCS Research)

Link to 2nd Ed of Course

About the Course: We will cover topics and recent research papers focusing on the broad goal of out-of-distribution generalization in machine-learning. Thus, in addition to the basic model, metric and optimization based meta-learning techniques, we will also cover topics and papers at the intersection of meta-learning and continual learning, disentangled representations, compositional and neuro-symbolic learning. We will apply these techniques to solve representative problems in few-shot, online and reinforcement learning and reasoning.

A major element of the course will be the course project wherein each student implements a paper from scratch, on data not included in the paper, and further works towards making a non-trivial improvement to it.

Lectures (live but online) will be as follows:

Part I: Introduction

Deep-learning refresher: MLP, RNNs, CNNs, Transformers
Basics of meta-learning: optimization-based meta-learning and MAML
Basics of meta-learning: model-based and metric-based meta-learning
Overview of project options and motivation for including these papers.
Representation learning and disentangled representations
Reasoning and Neuro-symbolic techniques, project finalization
Continual learning and Meta-reinforcement learning

Part 2: Recent Research

Recent papers in meta-learning
Recent papers in neuro-symbolic learning and reasoning
Recent papers in disentangled representations
Recent papers in meta-reinforcement learning
Recent papers in continual learning
Course summary, discussion on open research areas and final exam

Pre-requisite: Pytorch experience, and any course on deep-learning, or an equivalent online certification, or a self-statement of having gone through Yann Lecun's Deep Learning course at NYU - lectures 1.x, 2.x, 5.2, and 8.x available at: https://cds.nyu.edu/deep-learning/

Evaluation:

(i) 2 MCQ quizzes and 3 programming assignments given in class 10% and 20% (respectively)

(ii) Project involving implementing a paper from scratch on new data 40% and improving it 10%

(iii) Final Exam based on research papers covered in the course 20%

Time-frame: Jan - Apr 2022

TIMINGS: Tue and Fri 6-7:30pm see Dates for more details

Venue: Online: Zoom meeting invites will be shared on Slack workspace (invitation for the same will be shared on email-groups)

Google Sites

Report abuse