IPAB seminar - Moacir Ponti (USP)

Title:
Noise Robustness and Convergence of CNNs + Cross-domain Feature Embedding with Triplet Network

Abstract
This talk is divided into two parts. First, it presents an experimental investigation on the behaviour of Convolutional Neural Networks in the case of varying perturbations, in particular noise, in the training and test sets, providing insights about network overtraining and relationship between noise in the input data and the dropout strategy. Then, a study on the convergence of CNNs in the light of the Statistical Learning Theory is discussed, proposing an approach to estimate the Shattering coefficient of those CNN-based classification algorithms, which provides a lower bound for the complexity of their space of admissible functions (algorithm bias). In the second part, results on cross-domain feature embedding are presented, by using a multi-stage training approach and a triplet network, resulting in a model capable of performing feature embedding for the problem of sketch-based image retrieval across 33 different categories.

Bio:
Moacir Ponti received his PhD (2008) and MSc (2004) degrees at the Universidade Federal de São Carlos, Brazil. He is currently a Professor at the Institute of Mathematical and Computer Sciences, Universidade de São Paulo (USP), Brazil. He has been a Principal Investigator on research grants funded by Brazilian research agencies CNPq and FAPESP as well as international network projects, e.g. UGPN. He was recipient of the Latin America Research Award from Google in 2017. Author of more than 40 papers in peer reviewed journals and conferences, his research area is signal, image and video processing, with current focus in feature extraction and representation learning, with various applications.