Project: Convolutional Neural Networks for Auditory Attention in Robots

Key information:

Abstract: 

Deep learning has proven to be an immensely powerful tool, surpassing human performance in a wide range of fields such as agriculture, language processing and healthcare. One surprising, yet critical weakness of deep learning is a phenomenon known as ‘Catastrophic Forgetting’. This describes the inability of a neural network to learn something new without ‘forgetting’ its prior knowledge. Instead, classically, the neural network must be trained on all desired tasks and data simultaneously. In the real world this is often impossible or, at least, impractical. It means storing all incoming data permanently and retraining regularly, on this growing dataset. Human brains have inspired many advances in machine learning and offer proof that it is possible for a system to learning continuously over time, without requiring a continuously growing memory bank. The field of continual learning aims to combat Catastrophic Forgetting.

This project focuses on taking advances in continual learning into the field of anomaly detection, to help predict mechanical failure in robotics in a scalable and flexible way.