Poolability and Transferability in CNN. A Thrifty Approach

Abstract

The current trend in deep learning models for semantic segmentation are ever increasing model sizes. These large models need huge data-sets to be trained properly. However medical applications often offer only small data-sets available and require smaller models. A large part of these models' parameters is due to their multi-resolution approach for increasing the receptive field, i.e. alternating convolution and pooling layers for feature extraction. In this work an alternative parameter free approach is proposed to increase the receptive field. This significantly reduces the number of parameters needed in semantic segmentation models and allows them to be trained on smaller data-sets.

Tasks

Reproductions