Image Processing Lab (IPL)
Improved convolutional layers in deep learning via learned strides
Reference persons ENRICO MAGLI
Thesis type RESEARCH
Description A recent paper [R1] has shown that convolutional layers in deep neural networks can be improved by learning the stride parameter, instead of using a fixed stride. This is implemented as a cropping window having learnable size in the Fourier domain. This method simplifies the design of a deep architecture, and achieves performance gains in several tasks.
The objective of this thesis is to develop this concept even further, employing signal processing methods to optimize the cropping stage, and adopting signal-adaptive windows. The new methods developed during the thesis will be tested on classification problems, as well as other problems to be defined.
[R1] R. Riad, O. Teboul, D. Grangier, N. Zeghidour, "Learning strides in convolutional neural networks", Proc. of ICLR 2022, winner of best paper award.
Required skills Candidate students should have some background on neural networks. Some experience of TensorFlow environment and Python programming are desirable, along with good programming skills.
Deadline 11/05/2023 PROPONI LA TUA CANDIDATURA