Title: Opening the Black Box of Deep Learning: Information Flow and Condensation
Speaker: Naveensurya V (Masters thesis at NTU Singapore)
Date and Time: 16th October 2025 (Thursday) | 9 pm
Venue: PSB 1207
Abstract: This talk begins with a visual exploration of Convolutional Neural Networks (CNNs) layers, showing how raw images evolve into abstract feature maps as data moves deeper into the network. In contrast, the representations within fully connected networks (FCNs) remain largely a black box. We then uncover what drives these transformations using information theory and the Information Bottleneck (IB) principle. During training, neurons in each layer condense into clusters with similar outputs. This process reveals how neural networks naturally balance compression and relevance, shaping their remarkable ability to generalize.
Title: Epidemiological and Evolutionary Modelling
Speaker: Shreya Jugoolkar (Masters thesis at OIST-Japan)
Date and Time: 16th October 2025 (Thursday) | 9:30 pm
Venue: PSB 1207
Abstract: Mathematical modelling can serve as a strong framework for describing observable phenomena and translating qualitative observations into quantitative understanding. In epidemiology, models have been used for over a century to study the spread and control of infectious diseases, often offering insights regarding transmission patterns and public health strategies. Adaptive dynamics is a mathematical framework that emerged in the 1990s, used in evolutionary models. It allows exploration of the long-term consequences of small mutations in the traits expressing the phenotype. This talk will introduce the basic ideas and terminology used in both epidemiological and evolutionary models, and highlight how these perspectives can be integrated for a better understanding.