Thesis title: Artifacts mitigation in Artificial Neural Networks - Rethinking the building blocks with a Signal Processing approach
A precise understanding of the inner mechanism of Artificial Neural Networks is crucial for their usage in scientific research and their reliability as tools used in the everyday life. In the field of Neural Network Interpretability, Feature Visualization is a procedure used to try to understand what a single neural unit is searching for. Current Feature Visualization procedures do suffer of many artifacts and corruptions, and many tricks exist in literature to try to hide them.
In this thesis, we search the underlying causes of those artifacts. We find different contributions, such as aliasing arising from the violation of the Nyquist criterion in the donwsampling operations, aliasing from harmonics generation in the nonlinear operations and finally artifacts generated due to the sparsity induced in gradient backpropagation.
We then try to resolve them at their root, modifying the building blocks of Neural Network architectures in order to follow the best practices currently used in Signal Processing literature.
We start by defining a new method for downsampling Artificial Neural Networks feature maps, involving an explicit filtration in the frequency-domain. We name this method FourierPooling and we show its smoother action on gradient backpropagation and image formation in Feature Visualization.
We then formulate FundamentalAndFirstHarmonic: a nonlinear function with the minimal amount of possible aliasing, by limiting the harmonics it generates.
We also define the NormalizedCrossCorrelation, modifying the convolution operation in order to behave as a true similarity measure between the input and the kernel, allowing a clear interpretation of its inner mechanism.
Finally, we assemble all those building blocks to build a Neural Network which is able to keep aliasing under control.