Invited Speakers

Reinforcement learning is traditionally built around a simple and powerful abstraction: an agent acts to maximize a scalar reward. But is a single number always an adequate representation? In this talk, I will explore reinforcement learning beyond the assumption of a given scalar reward function, focusing on settings where objectives must instead be inferred from preferences and other, possibly diverse, forms of feedback. Starting from inverse and preference-based reinforcement learning, I will discuss the relationship between preferences, utility representations, and rewards, and ask when observed feedback can actually be explained by a scalar reward function. When this is not possible, richer representations become necessary: preferences may be incomplete, behaviors may be incomparable, and the underlying problem may be inherently multidimensional. I will then connect these ideas to RLHF, where human preferences are typically modeled through a learned scalar reward. This perspective highlights the representational assumptions underlying this common pipeline and suggests alternative ways of incorporating preferences and other forms of human feedback into reinforcement learning.

In spite of advances in vaccines and antibiotics, a tremendous amount of research effort has been invested in the recent years for discovering drugs for diseases without a known cure. Millions of people globally suffer from health conditions that necessitate the development of new and more effective treatments. The widespread and prolonged use of antibiotics has led to drug-resistant infections, making current treatments less effective. While antibiotic resistance is increasing rapidly, the development of new antibacterial drugs has slowed to an all-time low. This highlights the urgent need for novel methods to discover new antimicrobial agents. However, traditional drug development techniques are slow, inefficient, and expensive. To address these challenges, we have developed Deep Drug, an AI-powered drug design pipeline that can interpret vast datasets to identify potential new therapeutics. This technology can significantly accelerate drug discovery, reducing both development time and costs.

Climate change and global warming are today recognized as problems that threaten the very existence of humanity on earth. In the recent years, human civilization has been threatened by natural disasters – hurricanes, tornadoes, rise in sea level, droughts, floods, desertification of land, erosion of coastal areas, wildfires, etc. Human activity has resulted in water and air pollution that is adversely affecting population health. Encroachment of humans in animal habitat have threatened biodiversity and ecological balance and have increased the possibility of transmission of diseases from wildlife to humans leading to possible pandemics/epidemics. Environmental disasters have diminished agricultural productivity and led to proliferation of pests, lack of grazing land resulting in reduced animal husbandry, causing increased poverty and reduced food security.

We show how AI can help in tackling the environmental challenges mentioned above through information gathering, processing and analyzing, decision making, observing consequences, and feeding back for forecasting. We discuss our recent work in using AI/ML for carbon mapping, hypoxia predictions, climate smart agriculture, wildfire prediction and detection, preserving biodiversity, and energy efficiency in built environment.