on the May 7, 2021
Online
Program
10:30 - 10:45
Amar Bouali
Introduction to the Partnership and Innovation initiatives of 3IA
10:45 - 11:05
Antoine Collin (CNRS)
Automatic cell type annotation for cell atlas construction
Abstract: Chronic respiratory diseases (COPD, fibrosis, asthma, cystic fibrosis, ciliopathies) affect several hundred million people worldwide and the COVID pandemic has recently drawn world attention on the impact of infectious respiratory diseases. To better understand the importance of the different resident cell types in these processes and determine useful new targets for developing a therapeutical arsenal, we are developing the use of single-cell RNAseq analysis. We demonstrate cell composition of the human airways by capturing the gene expression profile at a single-cell resolution. Our work will contribute to the lung seed network of the Human Cell Atlas international consortium, which aims to build a cellular atlas of the whole human body. One of the most challenging step in building a cell atlas is to assign a cell type (Multiciliated, Secretory, Basal, Lymphocytes…) based on the gene expression profile. This is an iterative process that involves dimensionality reduction, clustering and differential expression analysis. It produces marker genes which are used by an expert biologist to assign a cell type to a cell cluster. We aim to automate this time consuming process with the help of deep learning. This faces many challenges, mainly caused by extensive biological variability between datasets. Our first focus was to find metrics to assess the quality of cell annotation. As a preliminary work, we created a specificity metric characterizing the specificity of marker genes based on its distribution in every cell types. This is a starting point in the development of our own deep learning model for automatic cell type annotation.
11:05 - 11:25
Edouard Balzin (CNRS)
Do neural networks have homology?
Abstract: Neural networks and various techniques of learning are actively studied in the applied mathematics community, and some pure mathematical takes on the subject often come from the realm of differential calculus and geometry. Wondering if "derived mathematics" has any use to comprehend machine learning, I will talk about a construction that associates to a neural network, trained in a supervised learning context, various homology groups, based on a measure of correlations between activation values. We will discuss the behaviour of these groups in the process of learning and outline the (very many) challenges that one faces trying to make sense out of it all.
11:25 - 11:45
Ziming Liu (Inria)
High-resolution Detection Network for Small Objects
Abstract: Small object detection is a very challenging yet practical vision task. With deep network-based methods, the contextual information of small objects may disappear when the network goes deeper. An intuitive solution to alleviate this issue is to increase the input resolution, however, it will aggravate the large variant of object scale and introduce unbearable computation cost. To leverage the benefits of high-resolution images without bringing up new problems, we propose a High-Resolution Detection Network (HRDNet) which takes multiple resolution inputs with multi-depth backbones. Meanwhile, we propose the Multi-Depth Image Pyramid Network (MD-IPN) and Multi-Scale Feature Pyramid Network (MS-FPN). The MD-IPN maintains multiple position information using multiple depth backbones. Specifically, high-resolution input will be fed into a shallow network to reserve more positional information and reduce computational costs, while low-resolution input will be fed into a deep network to extract more semantics. By extracting various features from high to low resolutions, the MD-IPN can improve the performance of small object detection and maintain the performance of middle and large objects. Additionally, MS-FPN is introduced to align and fuse multi-scale feature groups generated by MD-IPN to reduce the information imbalance. Extensive experiments are conducted on the COCO2017 and the typical small object dataset, VisDrone 2019. Notably, our HRDNet achieves the state-of-the-art on these two datasets with significant improvements on small objects.
11:45 - 12:00
Additional questions and discussions