Connectionists: Resolving the interpretability of Convolutional Neural Networks (CNN) and Graph Convolutional Neural Networks (GCNN)

[Danilo Mandic]

Dear All,

I would like to draw your atteniton to a couple of hot-off-the-press 
articles by Ljubisa Stankovic and Danilo Mandic, which address the 
interpretability of CNNs and GCNNs through matched filters - a common 
method for finding patterns of interest in e.g. radar or communications.

Such an approach allows for a systems science interpretation of the 
whole convolution-activation-pooling chain, and even lends itself to the 
matched filtering interpretation of backpropagation. More detail can be 
found from:

https://ieeexplore.ieee.org/document/10054202

L. Stankovic and D. P. Mandic, "Understanding the basis of graph 
convolutional neural networks via an intuitive matched filtering 
approach'', IEEE Signal Processing Magazine, vol. 40, no. 2, pp. 
155-165, 2023.

and

https://ieeexplore.ieee.org/document/10021677

L. Stankovic and D. P. Mandic, "Convolutional neural networks 
demystified: A matched filtering based approach'', IEEE Transactions on 
Systems, Man and Cybernetics, in print, 2023.

Best wishes,

Danilo
---
Prof Danilo P. Mandic, Imperial College London, UK
President of the International Neural Networks Society (INNS)
Distinuished Lecturer of the IEEE Computational Intelligence Society
Distinguished Lecturer of the IEEE Signal Processing Society