Topic：Interpretable Convolutional NNs and Graph CNNs: Role of Domain Knowledge
Speaker：Professor Danilo P. Mandic, Imperial College London, UK
Time: Friday, August 11, 9:00 – 11:30 a.m.
Venue: F706, New Main Building, Xueyuan Road Campus
Abstract:
The success of deep learning (DL) and convolutional neural networks (CNN) has also highlighted that NN-based analysis of signals and images of large sizes poses a considerable challenge, as the number of NN weights increases exponentially with data volume — the so called Curse of Dimensionality. In addition, the largely ad-hoc fashion of their development, albeit one reason for their rapid success, has also brought to light the intrinsic limitations of CNNs — in particular those related to their black box nature. To this end, we revisit the operation of CNNs from first principles and show that their key component — the convolutional layer — effectively performs matched filtering of its inputs with a set of templates (filters, kernels) of interest. This serves as a vehicle to establish a compact matched filtering perspective of the whole convolution-activation-pooling chain, which allows for a theoretically well founded and physically meaningful insight into the overall operation of CNNs. This is shown to help mitigate their interpretability and explainability issues, together with providing intuition for further developments and novel physically meaningful ways of their initialisation. Such an approach is next extended to Graph CNNs (GCNNs), which benefit from the universal function approximation property of NNs, pattern matching inherent to CNNs, and the ability of graphs to operate on nonlinear domains. GCNNs are revisited starting from the notion of a system on a graph, which serves to establish a matched-filtering interpretation of the whole convolution-activation-pooling chain within GCNNs, while inheriting the rigour and intuition from signal detection theory. This both sheds new light onto the otherwise black box approach to GCNNs and provides well-motivated and physically meaningful interpretation at every step of the operation and adaptation of GCNNs. It is our hope that the incorporation of domain knowledge, which is central to this approach, will help demystify CNNs and GCNNs, together with establishing a common language between the diverse communities working on Deep Learning and opening novel avenues for their further development.
About the Speaker:
Danilo P. Mandic is a Professor of Machine Intelligence with Imperial College London, UK, and has been working in the areas of statistical machine learning, big data analytics, graph signal processing, bioengineering, and financial modelling. He is a Fellow of the IEEE and a current President of the International Neural Networks Society (INNS). Dr Mandic is a Director of the Financial Machine Intelligence Lab at Imperial and has more than 500 publications in international journals and conferences, with four research monographs on Wiley, Springer and Now Publishers.
Dr Mandic is a 2019 recipient of the Dennis Gabor Award for "Outstanding Achievements in Neural Engineering", given by the International Neural Networks Society. He was a 2018 winner of the Best Paper Award in IEEE Signal Processing Magazine, and a 2021 winner of the Outstanding Paper Award in the International Conference on Acoustics, Speech and Signal Processing (ICASSP) series of conferences. Dr Mandic served in various roles in the Word Congress on Computational Intelligence (WCCI) and International Joint Conference on Neural Networks (IJCNN) series of conferences, and as an Associate Editor for IEEE Transactions on Neural Networks and Learning Systems. He has given about 70 Keynote and Tutorial lectures in international conferences and was appointed by the World University Service (WUS), as a Visiting Lecturer within the Brain Gain Program (BGP), in 2015. Dr Mandic is a 2014 recipient of President Award for Excellence in Postgraduate Supervision at Imperial College.
School of Electronic Information Engineering
