A colloquium will take place on Monday, October 16, 2023, in ESIEE
14:30- 15:30 – Welcome coffee room 4352 – Epi 4 3rd floor
Title : Building Convolutional Neural Networks under the Expert’s Control
Abstract: The success of a convolutional neural network (CNN) mainly depends on its feature extractor, named encoder. However, training deep models with backpropagation from scratch requires considerable human effort in data annotation and network adaptation, leaving several questions unanswered: What is the simplest model for a given problem? How can it be trained with minimum human effort? Can we explain its decisions ?
Such questions ask for alternative methodologies to build CNNs under the expert’s control. We have presented Feature Learning from Image Markers (FLIM) — a methodology to estimate the encoder’s filters from markers drawn by the expert on discriminative regions of a few representative images. This talk explains how to use FLIM for building such a convolutional encoder layer by layer for object detection. The process also explores a single-layer adaptive decoder to create the entire CNN without backpropagation. We show that FLIM can build flyweight CNNs with equivalent or superior performance to state-of-the-art object detectors, being thousands of times lighter and suitable for CPU execution.
Download the slides here