Title: The side benefits of knowing an algorithm

Abstract: I was trying to help a group of school students discover a beautiful theorem in combinatorics. The theorem says: in any sequence of k^2+1 distinct numbers, there is either an increasing subsequence of k+1 numbers or a decreasing subsequence of k+1 numbers. The students eventually figured this out.  They also figured out why k^2 instead of k^2+1 would not do: there are sequences of k^2 numbers that have neither an increasing nor a decreasing subsequence of k+1 numbers. (Try to build one for k=3.) Just when I thought I had told them everything that I could about the theorem, one of them asked how many such sequences with numbers 1,2, …, k^2 were there. I did not know. I did not want to google. I wanted to understand. I went back to an algorithm I had learnt years ago. It gave me the answer. I understood a little more. I want to tell you what I found, and why that made me happy.

Bio: Jaikumar Radhakrishnan is a theoretical computer scientist with research interests in complexity theory, randomness and computation, quantum information and computation, combinatorics, and information theory. Radhakrishnan obtained his BTech in Computer Science and Engineering from IIT Kharagpur in 1985, and his PhD in Computer Science from Rutgers University, NJ, USA, in 1991.  He joined the Tata Institute of Fundamental Research in 1991; in 2024 he moved to the International Centre for Theoretical Sciences (ICTS-TIFR), Bengaluru, where he is currently a Distinguished Professor.

Title: Multilingual and Multicultural Responsible AI

Abstract: This talk explores the principles and challenges of implementing responsible AI in diverse linguistic and cultural contexts. It highlights the complexities of addressing gender bias in non-English languages and showcases efforts to develop culturally sensitive methods for bias identification and mitigation. The talk looks into the role of synthetic datasets and evaluation methods in ensuring accurate and inclusive AI applications. It underscores the importance of participatory design and community-centric approaches in uncovering nuanced biases and enhancing the effectiveness of responsible AI principles.

Bio: Kalika Bali is a Sr. Principal Researcher at Microsoft Research Labs India, where she has dedicated nearly two decades to enhancing human-computer interactions through language technologies. Her focus lies in creating inclusive tech for a diverse range of languages and communities, especially those that are underrepresented. She is particularly interested in how Foundational Models like GPT can impact society, for better or worse. Her recent work navigates the crossroads of multilingual and multicultural AI. She was on the first (2023) TIME100 AI list for her continuing work on breaking down language barriers and fostering inclusivity in the AI sphere.

Title: Fair Allocation of Indivisible Resources

Abstract: Fair allocation of indivisible items is a central problem in computer science and economics. Be it the division of inherited property among heirs, assigning courses to lecturers, sharing household chores among family members, distributing public housing units, or allocating healthcare resources, one of the objectives is to divide the resources “fairly” among stakeholders. This talk will introduce the field of Fair Allocation and provide several possible research directions.

Title:   Automated Synthesis: Fusing Formal Methods and AI

Abstract: We increasingly entrust large parts of our daily lives to computer systems, which are becoming more complex. Developing scalable and trustworthy techniques for designing, developing, and verifying these systems is crucial. In this talk, we will focus on automated synthesis, a technique that uses formal specifications to automatically generate systems (such as functions, programs, or circuits) that provably satisfy the specified requirements.

The talk will overview recent approaches that combine advances in automated reasoning, knowledge compilation, and machine learning to address a wide variety of practical automated synthesis problems. Additionally, we will introduce a state-of-the-art functional synthesis algorithm that leverages artificial intelligence to provide an initial guess for the system and then uses formal methods to repair and verify this guess, resulting in a system that is correct by construction.

Short Bio: Priyanka Golia is an Assistant Professor in the Computer Science and Engineering Department at the Indian Institute of Technology Delhi. Prior to that, she was a faculty at the CISPA Helmholtz Center for Information Security, Germany. She completed her Ph.D. through a joint degree program between Indian Institute of Technology Kanpur, India, and National University of Singapore.  Her research interests lie at the intersection of formal methods and artificial intelligence. Specifically, her work has focused on designing scalable automated synthesis and testing techniques. Her research has received  a Best Paper Nomination at ICCAD-21, a Best Paper Candidate at DATE-23, and an invitation for special issues in FMSD.

Title : Reflections on a journey in logic and automata across academia and industry

Abstract : I’ll briefly present my research work and my transition through various facets of logic and automata.

I’ll share my thoughts on research in academia, industry, and the Ph.D. journey.

Bio : Saptarshi Sarkar received his Ph.D. from IIT Bombay in 2022 in the area of logic, automata, and algebra.

He is a recipient of the Naik and Rastogi Award for excellence in Ph.D. research for the year 2020-2022.

His publications have appeared in top conferences and journals like LICS, CONCUR, LMCS etc.

After Ph.D., he joined SiFive as a formal verification engineer. Currently he is a senior formal verification engineer in SiFive.

Title: A Mathematician’s Journey into Building Robust Cryptosystems

Abstract: In this talk, I will share my research journey, beginning as a mathematician and venturing into the realm of theoretical computer science, with a particular focus on cryptography. I will discuss my work on building robust cryptosystems by using techniques from different related areas like coding theory, pseudorandomness, information theory, and more. Along the way, I will highlight several intriguing problems, ranging from theoretical ones to industry-relevant ones. I will also offer insights from my experiences, with key takeaways for PhDs and early-stage postdocs.

Bio: Sruthi Sekar is currently an Assistant Professor in the Department of Computer Science and Engineering at the Indian Institute of Technology Bombay. Her research interests span both theoretical and applied cryptography, as well as theoretical computer science more broadly. Prior to this, she was a postdoctoral researcher at the University of California, Berkeley. She earned her Master’s and PhD degrees from the Indian Institute of Science, Bangalore, and her undergraduate degree from Lady Shri Ram College, Delhi University. Her thesis, titled “Near-optimal Non-malleable Codes and Leakage Resilient Secret Sharing Schemes,” received an honorable mention at the ACM India Doctoral Dissertation Award 2022. She is the recipient of the Prof. P. L. Bhatnagar Medal (gold medal for her Master’s degree) from IISc. Her research has appeared at top conferences in theoretical computer science, cryptography, and security, and has gained recognition through several awards, including the TCS Ph.D. fellowship, Prof. Venkatchalaienger Scholarship (from the Mathematics Department at IISc), the IDRBT Doctoral Colloquium Award 2021, and the Sandwich Scholarship (facilitated her visit to Technion University, Israel).

Title: Learning Language Processing Models with Limited Labeled Data through Efficient Supervision, Data Transformations, and Specialized Architectures.

Abstract:  High-quality, human-labeled data is essential for enhancing the utility of self-supervised deep learning models. For example, the Llama-3 language model required fine-tuning on 10 million human-annotated examples despite being pre-trained on 15 trillion tokens. However, scaling dataset annotation across various languages, tasks, and domains can be prohibitively expensive and time-consuming, particularly for tasks requiring specialized expertise, such as converting natural language questions into SQL queries. Consequently, learning efficiently with less human supervision is important for enabling more inclusive and diverse AI applications.

In this talk, I will present my research on learning language processing models with limited human supervision. I will begin by discussing our methods for collecting more effective human supervision under modest data annotation budgets. We will explore how enabling annotators to generalize their labeled examples and model-driven data-selection techniques can help us build more accurate models with reduced human effort.  Next, I will present our strategies for generating high-quality labeled data for low-resource languages and domains. Often, valuable labeled data already exists in similar yet distinct contexts.  We will discuss techniques for effectively transforming existing datasets with language models to provide useful supervision for new languages and domains. In the final part of this talk, I will demonstrate how integrating task-specific priors into model architectures can result in more efficient and accurate models using the same amount of labeled data. I will conclude with a discussion of my current and future research directions.

Bio: Abhijeet’s research focuses on machine learning applied to text, speech, and code. He is currently a Senior Researcher at Microsoft, working on language models for code and their applications in AI copilots. He earned his PhD from IIT Bombay, where his research on data-efficient natural language and speech processing received multiple awards, including runner-up for the IKDD Best Doctoral Dissertation Award and an honorable mention for the ACM India Doctoral Dissertation Award. In 2018, he was one of twelve global recipients of the Google PhD Fellowship in Machine Learning. Previously, he worked as a software engineer at Samsung Research Institute and earned his bachelor’s degree in Electronics from IIT Kharagpur.

Title: GPU Computing Meets Persistent Memory

Abstract: A broad spectrum of application software, including AI, graph analytics, and scientific computing, today rely on massive parallelism of Graphic Processing Units (GPUs) to cater to their primary computing needs. However, an important class of applications, those that must make the results of their computation persistent (durable), has largely been left out from effectively harnessing the computing capabilities of GPUs. Examples of such applications include persistent key-value stores, databases, etc., that form the backbone of many internet-scale services, including Facebook and Instagram. This limitation is rooted in the inability of traditional GPU-accelerated programs to access and manipulate data on the persistent media directly.

We observe that it is possible to bring easy and efficient access to persistent data to GPU’s massively parallel computing through innovations across the software and the hardware stack. In this talk, we will present the key insights that enabled this newly found capability for GPU-accelerated software. We will provide an overview of a novel system that puts together emerging persistent memory technologies and GPUs through a new software runtime and a new class of GPU-accelerated application software.

Bio: Arkaprava is an associate professor at the Indian Institute of Science (IISc), where he mentors bright young minds in the Computer Systems Lab who are passionate about building new computing systems. His research often straddles the boundary between system software and computer architecture. Arka’s recent research focuses on enhancing the reliability and efficiency of GPU-accelerated software and on creating novel use cases for GPU-accelerated computing. Arka received the SiFive Rising Star Faculty Award in 2021 and the Google India Research Awards in 2021 and 2023.

Title: Neural Networks for Graph Combinatorial Optimization: Challenges, Insights, and Future Directions

Abstract: Solving Combinatorial Optimization (CO) problems over graphs has applications in various domains, including manufacturing, bio-sciences, and transportation. These problems are often NP-hard, leading to the development of approximation algorithms or heuristics, which rely heavily on domain-specific knowledge and past experiences with similar problems. This approach demands significant time and effort to tailor solutions to specific problem contexts. Can we leverage neural networks to automate this process? This is the central question we have been studying for the past five years, and in this talk, I will discuss our findings, highlighting key challenges, insights, and our perspective on the way forward. As a case study, we will explore our work on learning graph edit distances to highlight the advantages and limitations of using neural networks for learning heuristics in combinatorial optimization problems.

Bio: Sayan Ranu is the Nick McKeown Chair Associate Professor at the Department of Computer Science & Engineering and the Yardi School of AI, IIT Delhi. His research focuses on machine learning and data mining for graphs. He has received several honors, including the INSA Associate Fellowship (2023), IIT Delhi’s Early Career Faculty Research Award (2023), and IASc Associate (2020). His works have been recognized among the top-10 papers at WSDM 2023 and ICDM 2016, received the Best Paper Award at WISE 2016, the Most Reproducible Paper Award at SIGMOD 2018, and the Audience Appreciation Award at KDD 2023. He holds five US patents, and his dissertation inspired a life sciences R&D startup in drug discovery.