Learning CNN filter resolution using multi-scale structured receptive fields

January 12, 2018 at 3:30 pm by

Affiliation: Assistant professor at the Technical University Delft

Place: Large Lecture Room

Abstract

The design of filter layers in a CNNs is a matter of trail and error where the filter-size in a single layer is typically hard-coded. Here I question this design. Instead of hard-coding we aim to learn the resolution. We do this by coupling resolution to the standard deviation of a Gaussian blur kernel, and then learn CNN filters by learning coefficients of a local differential Gaussian basis. Preliminary results show that global resolution can be learned by optimizing the standard deviation, and –in contrast to pixel filters CNNs– is robust to changing scales.

 

 

Picture: