Trainable Spectrally Initializable Matrix Transformations in Convolutional Neural Networks

Alberti, Michele; Botros, Angela; Schütz, Narayan; Ingold, Rolf; Liwicki, Marcus; Seuret, Mathias (5 May 2021). Trainable Spectrally Initializable Matrix Transformations in Convolutional Neural Networks. In: 2020 25th International Conference on Pattern Recognition (ICPR) (pp. 8204-8211). IEEE 10.1109/ICPR48806.2021.9412204

Text Alberti__2021_.pdf - Published Version Restricted to registered users only Available under License Publisher holds Copyright. Download (2MB)

In this work, we introduce a new architectural component to Neural Network (NN), i.e., trainable and spectrally initializable matrix transformations on feature maps. While previous literature has already demonstrated the possibility of adding static spectral transformations as feature processors, our focus is on more general trainable transforms. We study the transforms in various architectural configurations on four datasets of different nature: from medical (ColorectalHist, HAM10000) and natural (Flowers) images to historical documents (CB55). With rigorous experiments that control for the number of parameters and randomness, we show that networks utilizing the introduced matrix transformations outperform vanilla neural networks. The observed accuracy increases appreciably across all datasets. In addition, we show that the benefit of spectral initialization leads to significantly faster convergence, as opposed to randomly initialized matrix transformations. The transformations are implemented as auto-differentiable PyTorch modules that can be incorporated into any neural network architecture. The entire code base is open-source.

Interest & Impact

Downloads

3 since deposited on 25 Jun 2021
0 in the past 12 months

Citations

5 Citations in Web of Science ®
6 Citations in Scopus

Search

in Google Scholar™

Services

Actions (login required)

Edit item

Actions (login required)

Edit item