Self-Supervised Learning for Complex Visual Understanding Loughborough University

Overview

Key facts of the Self-Supervised Learning for Complex Visual Understanding programme offered by Loughborough University 

In the field of machine learning, traditional supervised learning requires human experts provide labelled examples and explicit instructions to train AI models. However, obtaining large labelled datasets can be expensive and time-consuming. Most importantly, SL seriously restricts the machine's capability to generalise and learn new unseen scenarios.

Inspired by human learning, SSL leverages the intrinsic structure within unlabelled data to train models effectively. Self-supervised learning is a machine learning paradigm where a model learns from the data without explicit external labels, creating a surrogate task that the model can learn to solve without the need for external annotations.

Example Research Components:

• Contrastive Learning: The model is trained to differentiate between positive pairs (similar samples) and negative pairs (dissimilar samples).

• Representation Learning: learn rich and meaningful representations from visual data without explicit supervision. The goal is to enable the model to automatically discover relevant features and hierarchical representations.
• Predictive Learning: The model is trained to predict certain parts of the input data. e.g. predicting the next word in a sentence, filling in missing parts of an image.
• Generative Models: Training a model to generate a part of the input data or the entire input from some part of it, e.g. generative adversarial networks (GANs).
• Multi-Modal Integration: Explore the integration of multiple modalities, such as images, text, and possibly other sensor data, to improve the model's understanding of complex scenes. This can involve developing cross-modal self-supervised learning techniques.
• Temporal Learning: Learning representations based on the temporal order of data