Abstract: Deep Reinforcement Learning has been very successful in solving a variety of hard problems. But many RL architectures treat the action as coming from an unordered set or from a bounded interval. It is often the case that the actions and policies have a non-trivial structure that can be exploited for more efficient learning. In this talk, I will present several scenarios in which taking advantage of the structure leads to more efficient learning. In particular, I will talk about some of our recent work on learning representations for actions that capture the underlying spatial structures and on learning ensemble policies.

Abstract: Deep neural networks have revolutionized computer vision with their impressive performance on vision tasks. Recently their object representations have been found to match well to the visual areas of the brain. Yet their performance is still worse than humans, and it has been challenging to derive insight into why deep networks work or how they can be improved. In our lab we have been comparing object representations in brains and deep networks with the aim of understanding how we see and to make machine see better. We have shown that systematic biases in deep networks can be identified by comparing with brain representations, and that fixing these biases can improve performance. We have also been testing deep networks for the presence or absence of a variety of classic perceptual phenomena. Taken together these results suggest that accumulated wisdom from vision neuroscience can help us understand and improve deep neural networks. For more information visit our research group, the Vision Lab at IISc at https://sites.google.com/site/visionlabiisc/

Abstract: The increasing capabilities of machine learning algorithms is enabling the research community to address a number of long standing computer vision problems. However, as the saying goes that beauty lies in the eyes of the beholder, a technology can be utilized for both positive and negative tasks. For instance, while face recognition can provide solutions to problems like missing children and injured face identification, it can also be misused for similar tasks. We will discuss research initiatives in face recognition and deepfake that we have been pursuing along with the technological contributions and the social impact in the community.

Title of Talk: Convex Optimization: Tools and Applications in Statistics and Data Science

Abstract: Optimization plays an important role in Statistics and Data Science and many algorithms rely on estimators that result from solving convex optimization problem. I will introduce "Disciplined Convex Optimization" (DCP), a constructive approach to formulating such problems. DCP provides mathematical building blocks with known properties along with a set of rules to combine them. Although these rules are sufficient (but not necessary) conditions for convexity, the approach captures a sizeable class of problems researchers encounter. Specifically, I will describe our work on CVXR, which implements DCP in the R programming language. To solve a convex problem, one specifies an objective and constraints by combining constants, variables, parameters, and a library of functions in a manner that closely mirrors the actual mathematical description. CVXR then applies DCP rules to verify the problem's convexity. Once verified, the problem is converted into standard conic form using graph implementations and passed to a numerical solver. If time permits, I will demonstrate with some examples.

Abstract: The field of health informatics encompasses a wide range of disciplines aimed at acquiring, processing, and interpreting data to improve human health and healthcare services. Within the healthcare domain, informatics encompasses several specialized fields, including cheminformatics, pharmacoinformatic, health informatics, and medical/clinical informatics, which collectively generate an extensive amount of biological and clinical data. Bioinformatics, in particular, plays a pivotal role in compiling, storing, annotating, and analysing this data, making it accessible to both biologists and non-biologists alike. In light of the escalating global healthcare burden caused by emerging infectious diseases, the need for efficient drug discovery and development has become paramount. Informatics advancements, such as cheminformatics and pharmacoinformatic, have significantly reduced costs and time associated with drug discovery by enabling the identification of drug targets, selection of lead compounds, and prediction of crucial drug properties. Moreover, the field of immunoinformatic has emerged as a crucial player in vaccine development, employing computational tools to expedite the discovery of potential vaccine candidates. Medical/clinical informatics deals primarily with patient data, clinical knowledge, and information related to patient care, playing a pivotal role in disease diagnosis through the identification of biomarkers and assisting healthcare professionals in providing personalized treatments. The advent of the Internet of Things (IoT) has further revolutionized healthcare by facilitating the development of mobile apps, telemedicine platforms, and wearable sensor-based devices that enable remote monitoring and real-time health data collection. This talk aims to provide a comprehensive overview of freely available computational tools and databases in key areas of healthcare, such as drug discovery, toxicity and adverse effects assessment, vaccine development, disease diagnosis, and IoT applications. The featured resources encompass databases, web servers, standalone applications, and mobile apps, offering a diverse range of support to researchers across various healthcare disciplines. By highlighting these valuable resources, researchers can leverage their functionalities to expedite their work and contribute to the advancement of human health and healthcare services (https://webs.iiitd.edu.in/ )