On the 2024-06-20, at 1:30 PM, there will be two talks in Room 1201@ESIEE
First talk:
Quantile Based Approaches for Robust Classification
A. Challa
Abstract: Robustness to distortions is a key challenge affecting the reliability of machine learning models. This talk explores using quantile-based approaches to address this issue. We focus on two main questions — (1) How can we make any given classifier return well-calibrated probabilities without compromising its properties? (2) How can we build networks that incorporate robustness within the training process itself? We find that quantiles offer a simple and elegant solution to both problems. Quantiles have an intrinsic property of characterising the underlying distribution and hence can be used for efficient control of ML models. In simple words, we address (1) by forcing the network to learn the output quantiles, and (2) by using quantiles to normalize training across distributions.
Brief Bio: Dr. Aditya Challa has obtained his bachelor’s degree from Indian Statistical Institute, masters from University of Warwick. He completed his PhD from Indian Statistical Institute under the guidance of Prof. B.S.Daya Sagar and Prof. Laurent Najman (as a co-supervisor). He is currently working as an assistant professor at Birla Institute of Technology and Sciences, Goa Campus. He is primarily interested in the fields of mathematical morphology, machine learning and its statistical underpinnings.
Second talk:
Watershed as Greedy 1NN: Regularising 1NN classifier for Improved Generalization
S. Danda
Abstract: In this talk, I will present a novel approach to deep learning classification that challenges the dominance of linear classifiers. Traditionally, a deep learning classifier is represented as a composition of two parameterized functions g o f, where f learns the embedding and g classifies the embeddings. Despite the success of linear classifiers, non-parametric methods like Nearest Neighbour classifiers have not performed as well. Our proposed method introduces a non-parametric approach that matches the performance of linear classifiers. We demonstrate that the hyperparameter ‘K’ in KNN acts as a regularization mechanism, but is not optimal. We then propose a new regularization technique, the watershed classifier, which employs a greedy approach. By introducing a loss function tailored to the watershed classifier, we achieve superior performance compared to Neighbourhood Component Analysis (NCA). Our results show that the non-parametric watershed classifier performs on par with, or better than, traditional linear classifiers trained with cross-entropy.
Brief Bio: Dr. Sravan Danda has obtained his bachelor’s degree in Mathematics and master’s degree in Statistics, both from Indian Statistical Institute. He completed his PhD from Indian Statistical Institute under the joint guidance of Prof. B.S.Daya Sagar and Prof. Laurent Najman, He is currently working as an assistant professor at Birla Institute of Technology and Sciences, Goa Campus. His current research interests include applications of machine learning for agriculture, mathematical morphology and robust machine learning.
Localisation
Salle 1201 (ESIEE Paris)