Profile
Project title:
Information-theoretic neural networks for online perception of auditory objects
Abstract:
The power of deep neural networks is well evidenced in producing effective embeddings of complex data. However, the theory of deep learning is not well understood and shows a severe reliance on training data, casting a shadow over the robustness and generalisation of deep networks. A way forward is to merge neural networks with information-theoretic learning processes: if applied to sensor data (e.g. image, video, audio), this can furthermore yield organised methods of self-supervised learning that mimic cognitive learning in humans. This is particularly relevant to audio models and representations.
C4DM theme affiliation:
Machine Listening, Music Cognition