Advancing Continual Learning for Robotics Applications

This lecture explains the mathematics behind a simple multi-layer perceptron with an input layer, one hidden layer, and an output layer.

This lecture covers how convolutional filters extract feature maps from images and how recurrent neural networks (RNNs) handle temporal or time-series data.

This tutorial covers the implementation of a MLP network, a CNN network and a RNN network on some dummy data

This lecture explores the limitations of traditional ML in non-iid (independent and identically distributed) scenarios, where data distribution shifts over time, leading to the catastrophic forgetting problem.

This lecture introduces the concepts of continual learning, also called lifelong or incremental learning, and discusses its advantages.

This lecture explores realistic continual learning scenarios using the MNIST dataset in class-incremental and task-incremental settings.

This lecture further extends the concepts by explaining in detail how experiences are generated, and the difference between the different continual learning scenarios and when to use each.

In this tutorial, we show how traditional DL algorithms suffer from catastrophic forgetting when trained on data incrementally.

In this lecture we discuss about the various metrics that can be used to evaluate the performance of the continual learning algorithms.

The lecture introduces the various continual learning strategies present in the literature with an in depth discussion about the regularization methods.

This lecture builds on the previous one and explains the intuition behind the other methods namely architectural, rehearsal, generative replay, and hybrid strategies.

The tutorial covers the implementation of a Continual Learning algorithm using the Avalanche framework

This lecture covers the rehearsal based CL approaches as well as the hybrid methods to address catatstrophic forgetting in the network

In this final lecture, we explore a real-world application of a Continual Learning algorithm on a soft robotic arm and conclude with final remarks on CL advancements.