Training neural networks 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 experts control the training process? This lecture presents ongoing research towards creating convolutional neural networks (CNNs) for object detection, segmentation, and identification using very few representative images. Its results benefit diagnostic medicine, in which data annotation is costly and sometimes impractical, and the diagnosis of gastrointestinal parasites is taken as an example. Feature extraction is a crucial stage performed by the CNN’s encoder. One can append a decoder for object detection, a classifier for object identification, or a decoder followed by a delineator for object segmentation. The talk shows how experts can control feature extraction, such that the encoder’s parameters are estimated from a few markers (weak supervision) placed on discriminative image regions. The talk then introduces an adaptive decoder followed by a delineator for object segmentation, demonstrating how to create CNNs with minimum human effort and no need for backpropagation. The current results for the segmentation and identification of gastrointestinal parasites are presented. Finally, the concluding remarks provide directions for other applications, representative image selection, and classifier training with reduced human effort in data annotation.
Bio: Alexandre Xavier Falcão is a Professor in Computer Science at the Institute of Computing, State University of Campinas (UNICAMP). He holds a PhD from UNICAMP (1997), focusing on medical image analysis at the University of Pennsylvania from 1994-1996. He has been in the image analysis field for over 31 years, with projects in video quality assessment (Globo TV, 1997), plant phenotyping (Cornell University, 2011-2012), and several other image analysis applications developed at UNICAMP since 1998. He has authored over 360 papers and licensed over ten technologies, with five currently in the market. His research interests cover image analysis, data visualization, and human-machine interaction by combining humans’ superior cognitive abilities with machines’ higher data processing capacity.
Localisation
Amphithéâtre 163 (ESIEE Paris)