Abstract: We study the error rate of LLMs on tasks like arithmetic that require a deterministic output, and repetitive processing of tokens drawn from a small set of alternatives. By analyzing the accumulation of errors in the attention mechanism, we theoretically derive a quantitative two-parameter relationship between the accuracy and the complexity of the task. We empirically verify our formula across a range of tasks and state-of-the art LLMs find excellent agreement between the predicted and observed accuracy in many cases. We also identify deviations in some cases that lead us to interesting insights about the functioning of models. We show how this understanding helps to construct prompts to reduce the error rate.

About the speaker: Suvrat Raju is an Indian physicist whose research focuses on quantum gravity and quantum field theory. Suvrat has worked on black holes, focusing on the ‘information paradox’ around them. He has formulated the Papadodimas-Raju proposal for black holes. Suvrat studied physics at St. Stephen’s college in Delhi University and went on to complete his PhD at the Harvard University. He is currently a professor at the International Centre for Theoretical Sciences of the Tata Institute of Fundamental Research (TIFR).

Abstract: Despite rapid progress in the performance of State-of-the-Art AI models for quantum, most such methods remain blackboxes: lacking guarantees of robust and reliable predictions that meet uncertainty quantification benchmarks essential in scientific domains. To address this gap, I will summarize a few directions that improve robustness, mechanistic interpretability, and uncertainty quantification of complex learning and sample generation abilities. I will present the simplest model capable of in-context learning, an ability that underpins LLM success, especially for quantum attributes. The emergence of in-context learning ability is studied for the simplest class of tasks: linear regression. The model performance is exactly derived in the joint asymptotic limit of a large number of samples, token dimensions, sample length, and task diversity: exhibiting neural scaling laws and a phase transition from memorization-to-generalization. These results are supported by experiments on standard (full) architectures. Time permitting, I will also introduce a framework hinging on “Renormalization Group”, a cornerstone of theoretical physics, that systematically coarsegrains data features irrelevant to learning, while capturing nontrivial perturbations to model predictions within scientific uncertainty bounds. Altogether, these works advance AI reliability in a first-principles manner, while bridging AI with fundamental physics.

About the speaker: Dr. Anindita Maiti is a Postdoctoral Fellow at the Perimeter Institute for Theoretical Physics, cross-affiliated with Prof. Roger Melko’s group Perimeter Institute Quantum Intelligence Lab (PIQuIL) since September 2023. Previously, Anindita held a short postdoctoral appointment in physics for ML foundations, mentored by Prof. Cengiz Pehlevan, at Harvard Applied Math (May–August 2023). She earned her PhD in theoretical high-energy physics (string theory and particle theory division) at Northeastern University and the NSF AI Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) in May 2023, under the supervision of Prof. James Halverson. Anindita received her Integrated Bachelors and Masters in Engineering Physics at IIT Bombay in Aug 2017, guided by Prof. Urjit Yajnik. Anindita’s research lies at the intersection of AI/ML, quantum, and statistical physics. In short, she works on Physics of Learning and ML for Quantum. Broadly, Anindita uses theoretical physics concepts: such as path integrals, renormalization group flow, computational statistics, and random matrix theory, to develop a physics-informed theoretical foundation for ML and to guide the principled design of AI systems. Anindita applies this framework to construct interpretable and trustworthy AI- and ML-based simulation strategies for quantum field theory and quantum many-body physics.

Abstract: I will start with a few examples where confounders disrupt classical inference paradigm and motivate why a causal approach is needed. I will review the basic formalism and central ideas of causal inference such as Neyman’s potential outcomes (counterfactuals) and Don Rubin’s foundational assumptions under which causal inference is possible.

Abstract: The continental-scale reorganization of atmospheric circulation that accompanies the onset of India’s summer monsoon has been studied for over a century, and the progressive spatial expansion of rainfall that occurs during this transition has great consequence for more than a billion people. Smallholder farmers in India make consequential decisions about planting based on their expectations of when this transition will occur, and they benefit from multiple weeks of advance notice of the onset date. However, no weeks-ahead forecast of the timing of this rainfall transition in regions across India has been available to farmers or shown to be skillful. Here, we describe the development and deployment of a new forecasting system for regional, agriculturally relevant monsoon onset that blends artificial intelligence (AI) weather prediction models with a new statistical model for the probability of first-occurrence events in a season. A novel feature of this statistical-AI system is its ability to account for the likelihood of dry periods beyond the lead-time horizon of the AI model output. We describe the methods used to deploy this AI-based system in real time, and discuss the successful dissemination of probabilistic onset date forecasts by SMS to 38 million farmers in the summer of 2025.

About the speaker: William Boos is a professor at the University of California, Berkeley in the Department of Earth and Planetary Science, where he specializes in atmospheric dynamics. His research focuses on understanding the mechanisms that govern extreme weather phenomena, planetary wind patterns, and monsoons, which are continental-scale atmospheric circulations that deliver water to billions of people in Earth’s tropics. He is also a Faculty Scientist at the Lawrence Berkeley National Laboratory. Before moving to Berkeley, he was on the faculty at Yale and a postdoctoral fellow at Harvard. He received his Ph.D. from MIT and his undergraduate degrees from the State University of New York at Binghamton.

Website:

Abstract: Artificial intelligence is beginning to reshape work at a speed few institutions are prepared for. Tasks that once required years of training can now be performed, assisted, or accelerated by machines that learn and improve continuously. As this capability spreads, the foundations of how organizations operate and how individuals create value are starting to shift. The question is no longer whether AI will change work, but how deeply it will transform the roles we play, the skills we develop, and the systems we build around human potential.

Short Courses Announcement

Abstract: As AI systems become integral to knowledge work, the future of work and learning is increasingly hybrid. In this talk, I will discuss a modeling perspective on human–AI interaction, drawing on my recent work on AI-assisted labor, and, time permitting, examine how delegation and feedback may reshape how people work and learn.

Speaker Bio

Devansh Jalota is a PhD candidate in Computational and Mathematical Engineering at Stanford University, where he is a Stanford Interdisciplinary Graduate Fellow. His research develops data-driven learning algorithms and incentive schemes to advance the science and practice of market design for sustainable resource allocation, with a particular focus on applications in future mobility systems and electricity markets. Prior to joining Stanford, he received his bachelor’s in applied mathematics and civil engineering at UC Berkeley.

Abstract

Technological advances have opened new avenues for designing market mechanisms for resource allocation, from enhancing resource allocation efficiency with widespread data availability to enabling real-time algorithm implementation. While these technological advancements hold significant promise, they also introduce new societal challenges pertaining to equity, privacy, data uncertainty, and security that existing market mechanisms often fail to address. My research develops data-driven and online learning algorithms and incentive schemes to address these challenges of traditional market mechanisms, thereby advancing the science and practice of market design for sustainable and society-aware resource allocation.

In this talk, I focus on addressing data uncertainty and privacy issues in the context of Fisher markets, a classical framework for fair resource allocation where the problem of computing equilibrium prices relies on complete information of user attributes, which are typically unavailable in practice. Motivated by this practical limitation, we study a modified online incomplete information variant of Fisher markets, where users with privately known utility and budget parameters, drawn i.i.d. from a distribution, arrive sequentially. In this novel market, we establish the limitations of static pricing and design dynamic posted-price algorithms with improved guarantees. Our main result is a posted-price algorithm that solely relies on revealed preference (RP) feedback, i.e., observations of user consumption, achieving the best-known guarantees for first-order algorithms in the RP setting while providing a regret analysis of a fairness-promoting logarithmic objective, unlike typical non- negative and bounded eDiciency-promoting objectives in online learning.

Sarvesh Ravichandran Iyer is a post doctoral student under Prof. Sandeep Juneja and a member of the SCDLDS since May 2025. He obtained his dual BS-MS degree from IISc in 2018, and his PhD. in 2024 from ISI, Bangalore. He was a visiting faculty in the mathematics department at Ashoka in the academic year 2024-25. His research interests revolve around pure jump Levy processes and their applications.

Abstract: Exponential tilting is a technique in rare event sampling that uses an underlying change of measure in a probability space, making these rare events more likely. It is applied in multiple domains like finance and climate science to aid with the prediction and mitigation of extreme events. Typically, given some number of samples from an unknown distribution, one is required to produce a large number of exponentially twisted samples from these in order to predict the likelihood of such events. We achieve this aim in two steps. First, we reweigh the samples of the original distribution appropriately using a twisted empirical estimator, and subsequently perform diffusion sampling on the output of the estimator, thereby obtaining more twisted samples. We delineate regimes where the empirical estimator performs well, and where it does not : informally, large twists cannot be performed with very few samples, and twisting unbounded distributions is harder than twisting bounded distributions. We also provide theoretical guarantees on the accuracy of diffusion sampling in these regimes.

Riddhipratim Basu got his bachelor's and master's degrees from Indian Statistical Institute, Kolkata and his Ph.D. from the Statistics department of University of California, Berkeley. After a postdoctoral stint at Stanford he joined the International Centre for Theoretical Sciences of Tata Institute of Fundamental Research where he is currently a faculty. He is interested in various aspects of probability theory and its applications including models of random growth, random matrices and interacting particle systems.

Abstract: Instances of stochastic growth is ubiquitous in nature, examples include spreading of rumour, infection or forest fires among many others. Various mathematical models have been designed to explain large scale behaviour of such systems and have been extensively studied both in statistical physics and in probability theory over the last few decades. I shall describe some of these models, focussing primarily in the planar case, which are conjectured to exhibit universal features of the so called Kardar-Parisi-Zhang (KPZ) universality class. I shall describe some of the major conjectures in the area and discuss a bit of what is known rigorously.

Amit Deshpande studied mathematics and theoretical computer science at Chennai Mathematical Institute (CMI) and Massachusetts Institute of Technology (MIT). He joined Microsoft Research India after graduation and continues to work there. His current interests lie in responsible AI, more specifically, in fixing data bias and plumbing AI/ML pipelines.

Abstract: AI/ML systems routinely classify and rank data, products, media, even people. Allowing randomized or stochastic decisions is even more contentious. On the one hand, we toss a coin to break ties fairly all the time. On the other hand, even a fair coin toss discriminates between similar individuals for no valid reason. In this talk, we will debate some provable benefits and drawbacks of randomization in fair classification and ranking. For example, can a randomized fair classifier give better accuracy than a deterministic fair classifier? Is a randomized fair classifier more robust to data shift than a deterministic fair classifier? Can a randomized ranker maximize ranking utility while allowing fair representation of two demographic groups in the top ranks, not just in expectation but in every output?

Parthe Pandit is a Thakur Family Chair Assistant Professor at the Center for Machine Intelligence and Data Science (C-MInDS) at IIT Bombay. He was a Simons postdoctoral fellow at UC San Diego. He obtained his PhD from UCLA, and his undergraduate education from IIT Bombay. He has received the AI2050 Early Career Fellowship from Schmidt Sciences in 2024, and the Jack K Wolf Student Paper Award at ISIT 2019.

Manmeet Singh is presently at the Jackson School of Geosciences, The University of Texas at Austin, Austin, USA. He served as a Staff Scientist at the nodal national lab on weather and climate in India, Indian Institute of Tropical Meteorology, Ministry of Earth Sciences, Govt of India for 11 years from 2013-2024. He was also a Fulbright-Kalam fellow at the Jackson School of Geosciences, The University of Texas at Austin in 2021. His research interests include climate solutions to the problems on land, ocean and atmosphere using mathematical models, particularly numerical weather prediction systems. He is especially interested in AI/ML techniques, causal approaches, recurrence plots, complex networks and non-linear time series analysis for solving grand challenges in Earth System Science. He is an experienced climate modeller having contributed to the IITM Earth System Model simulations towards the IPCC AR6 report. Together with his PhD co-advisor, he developed and coupled the aerosol module of the IITM Earth System Model. He is active in teaching and has given invited talks at venues such as the NASA/UAH Seminar series, Microsoft India podcast among others. His PhD focussed on the impacts of the proposals suggesting volcanic eruptions as an analogue of solar geoengineering to halt climate change. Recently, his work has shown substantial improvements in high-impact short-range numerical weather predictions using deep learning and he has also developed novel physics inspired deep learning algorithms for high-resolution downscaling.

Abstract: In recent years, the confluence of data science, machine learning, and decision analytics has opened transformative pathways in understanding and addressing complex problems in weather and climate. This talk will explore how these interdisciplinary approaches are reshaping forecasting, risk assessment, and climate resilience planning. Drawing on recent case studies and ongoing research, we will examine the role of data-driven models in improving prediction accuracy, the integration of learning algorithms in climate simulations, and the use of decision frameworks to support adaptive responses to extreme weather events. The session aims to highlight both the opportunities and challenges in operationalizing these methods for real-world impact in a rapidly changing climate landscape.

Michael Gordy is a 30-year veteran of the Federal Reserve Board in Washington DC. His current research focus is econometrics, specifically in backtesting of forecast distributions and in machine learning. At the Fed, Michael serves as chief of a research unit that provides expertise on systemic risks in derivative markets, stress testing of dealer trading books, Treasury market reform, and market liquidity. He has held visiting appointments at Princeton and at Indian School of Business, and has served as co-Editor-in-Chief of the Journal of Credit Risk and as an associate editor of the Journal of Banking and Finance and the International Journal of Central Banking. In recognition of his contributions to the Basel II Capital Accord, Michael received Risk Magazine's 2004 Quant of the Year and GARP's 2003 Financial Risk Manager of the Year awards. Michael received his PhD in Economics from MIT in 1994.

Abstract: In this talk, we begin with a broad overview of risk management in the trading operations of large banks. At the heart of these management systems lies a forecast model for the distribution of profit and loss (P&L) over the next trading day. The empirical validity of this model is assessed by backtesting of the model's out-of-sample historical performance. Such backtests play a key role in supervisory examination of trading operations and in setting capital requirements.

We study a class of backtests in which the test statistic depends on a spectral transformation that weights exceedance events by a function of the modeled probability level. The weighting scheme is specified by a kernel measure which makes explicit the user’s priorities for model performance. We show how the class embeds a wide variety of backtests in the existing literature, and further propose novel variants which are easily implemented, well-sized and have good power. This framework is extended to allow general Lebesgue-Stieltjes kernel measures with unbounded distribution functions, which brings powerful new tests into the spectral class. Moreover, by considering uniform distribution preserving transformations of PIT values the test framework is generalized to allow tests that are focused on both tails of the forecast distribution.

Manik Varma is a Distinguished Scientist and Vice President at Microsoft Research India and an Adjunct Professor at the Indian Institute of Technology Delhi. He is best known for having started the research area of extreme classification in large-scale machine learning. Manik’s algorithms have made billions of predictions a day, have generated billions of dollars in revenue for businesses worldwide and have improved the productivity and experience of people in over a hundred and ninety countries. Manik is also known for his research on developing tiny classifiers that can fit within 2–16 KB of RAM and run on microcontrollers smaller than a grain of rice. His classifiers have been deployed on hundreds of millions of devices and have protected them from viruses and malware. Manik has served as an Associate Editor-in-Chief of the IEEE Transactions on Pattern Analysis & Machine Intelligence as well as a senior area chair at most of the premiere conferences in machine learning, artificial intelligence and computer vision. He is a Fellow of the Indian academies of science (IASc, INSA & NASI), the Indian National Academy of Engineering (INAE) and the Association for Computing Machinery (ACM). He is a recipient of the Bhatnagar Prize, has been a Visiting Miller Professor at the University of California at Berkeley and a Rhodes Scholar at Oxford.

Abstract:

Large Language Models (LLMs), such as GPT4 and O1, have delivered game-changing reasoning and synthesis capabilities leading pundits to proclaim that we have entered a new age of LLM reasoning. Yet, LLMs can make mistakes while answering simple factual queries, such as “Who did 51²è¹Ý hire most recently?”, and can fail spectacularly at complex tasks such as “Write a 2 page document summarizing the history of 51²è¹Ý”. A peek under the hood reveals that these mistakes are often caused by failures of the retrieval model which is responsible for fetching relevant information from the web and private databases. Such retrieval failures can lead to particularly egregious responses when the information required for completing the task at hand is not present in the LLM itself and needs to be fetched from external sources.

In this talk, I will discuss what the architecture and flow might look like for a retrieval platform for large-scale AI workloads that can significantly reduce such retrieval failures and thereby lead to much better LLM responses. I will also discuss how we can build a state-of-the-art generative retrieval model that forms the core of the platform and which can accurately fetch documents in milliseconds in a cost-effective manner. Finally, I will share some empirical evidence on how such a retrieval model might benefit millions of users around the world. Most parts of my talk should be broadly accessible to a lay audience.

Arun Sai Suggala is a Senior Research Scientist at Google DeepMind. He is broadly interested in optimization, online learning, reinforcement learning, and their applications to healthcare and AI for social good. His work has received the best student paper award at ALT'20. Arun obtained his PhD in Machine Learning from Carnegie Mellon University (CMU), working with Pradeep Ravikumar. Prior to CMU, he completed his undergraduate studies in Computer Science and Engineering in the Indian Institute of Technology, Bombay.

Short Courses Announcement

Schedule:

• February 21, 2025:

  • Jaikumar Radhakrishnan: 10:00 AM - 11:00 AM, 11:30 AM to 12:30 PM (Quantum Computation)

  • Piyush Srivastava: 2:30 PM-3:30 PM, 4:00 PM-5:00 PM (Causal Inference)

• February 22, 2025:

  • Piyush Srivastava: 10:00 AM - 11:00 AM, 11:30 AM to 12:30 PM (Causal Inference)

  • Jaikumar Radhakrishnan: 2:30 PM-3:30 PM, 4:00 PM-5:00 PM (Quantum Computation)

click here to download the schedule

Shravan Hanasoge is a Professor at the Department of Astronomy and Astrophysics at the Tata Institute of Fundamental Research and a co-Principal Investigator of the Center for Astrophysics and Space Science at New York University Abu Dhabi. He received his PhD and M.S. from Stanford University, B.Tech. from IIT Madras and has been a postdoctoral scholar at Princeton University, Courant Institute of Mathematical Sciences (New York University) and the Max-Planck Institute for Solar System Research.

Abstract: Machine learning can dramatically speed up astrophysical analyses, especially in the modern era of high-quality, large-scale observations. To advocate for the case of deep learning applied to fundamental science, I will describe results that we have obtained from asteroseismic analyses of the Kepler field. Stellar pulsations offer valuable insights into the internal structure and rotation profiles of stars. The availability of high-quality observations from numerous space-based instruments makes it possible to pursue ensemble analyses on an unprecedented scale. To this end, we have used machine learning to accelerate these studies by several orders of magnitude.

This endeavor presents unique challenges and opportunities for machine learning researchers. By collaborating with us, machine learning experts can contribute to advancing algorithms in areas such as unsupervised learning for pattern discovery in high-dimensional, unlabeled data, and developing robust models capable of handling noisy and irregular datasets characteristic of stellar spectra. Additionally, there is significant potential for innovation in model interpretability techniques tailored to complex scientific data, which can enhance transparency and trust in AI systems across various domains. This collaboration not only aids in solving complex astrophysical problems but can also drive methodological advancements in machine learning.

Arpit Agarwal is an Assistant Professor in the Department of Computer Science and Engineering at IIT Bombay. Prior to joining IIT Bombay, he was a postdoctoral researcher at FAIR Labs (Meta), where he worked with Max Nickel on socially responsible recommendation systems. Before that, he was a postdoctoral fellow at the Data Science Institute at Columbia University, hosted by Prof. Yash Kanoria and Prof. Tim Roughgarden. He completed his PhD in the Department of Computer & Information Science at the University of Pennsylvania, under the guidance of Prof. Shivani Agarwal. Arpit's research lies in the field of machine learning (ML) and artificial intelligence (AI), with a specific focus on the interaction between humans and ML/AI systems. His work spans topics such as learning from implicit, strategic, and heterogeneous human feedback; understanding the dynamics of human-AI interactions and their long-term influences; and designing AI systems responsibly to mitigate undesired consequences for individuals and society.

Abstract: Stochastic optimization is a widely used approach for optimization under uncertainty, where uncertain input parameters are modeled by random variables. Exact or approximation algorithms have been obtained for several fundamental problems in this area. However, a significant limitation of this approach is that it requires full knowledge of the underlying probability distributions. Can we still get good (approximation) algorithms if these distributions are unknown, and the algorithm needs to learn them through repeated interactions? In this paper, we resolve this question for a large class of “monotone” stochastic problems, by providing a generic online learning algorithm with √ T log T regret relative to the best approximation algorithm (under known distributions). Importantly, our online algorithm works in a semi-bandit setting, where in each period, the algorithm only observes samples from the r.v.s that were actually probed. Our framework applies to several fundamental problems in stochastic optimization such as prophet inequality, Pandora’s box, stochastic knapsack, stochastic matchings and stochastic submodular optimization.

There is a crisis of traffic congestion in Indian cities. Congestion slows down economic activity, and wastes time and energy. In the coming decades, as our cities become more affluent, more people will use more and bigger cars for their daily commute. Cars provide greater comfort and safety compared to two-wheelers, but also exacerbate the congestion problem. This problem needs to be addressed urgently.

In this talk, we will consider two approaches to reduce congestion: product innovation and policy intervention. First, we will discuss an innovative electric microcar called Robin that has the length and width of a motorbike, but provides the safety and comfort of a small car. We will discuss its design features and tradeoffs, and how it could reduce congestion by being a substitute for a car.

Second, we will propose a novel congestion pricing scheme that accounts for vehicle size and thereby incentivizes the use of smaller vehicles. This scheme can be used to charge vehicles for both driving and parking within the city. We will discuss a few variants of the basic scheme, and list several open questions to investigate.

This talk will illustrate how a combination of technological progress, product innovation and policy intervention can help solve serious societal problems.

Zoom Link:

Date: Tuesday, November 12, 2024
Time: 1:30 PM IST
Venue: AC01-LR-106
For details: ashoka-cdlds@ashoka.edu.in
or call: +91-9136857558

Organising committee: Vivek Borkar (IIT Bombay), Sandeep Juneja (51²è¹Ý), Praneeth Netrapalli (Google DeepMind), Sumegh P (51²è¹Ý), KS Mallikarjuna Rao (IIT Bombay)

Dr. Vivek Raghavan is an entrepreneur, technologist and creator of Digital Public Goods (DPG). He has embarked on a new mission as the founder of Sarvam AI to develop full stack Generative AI for India.

As Chief AI Evangelist at the EkStep foundation, Vivek was an advisor to Digital India Bhashini (National Language Translation Mission) which aims to make available all services and information to citizens in their local language through the medium of voice. Vivek was also Chief Mentor for the Nilekani Center at AI4Bharat, IIT Madras whose aim is to achieve English equivalence in language AI for all Indian languages through open datasets, models and applications.

Vivek was the Chief Product Manager and Biometric Architect at the Unique Identification Authority of India (UIDAI). He has been responsible for the design, implementation and scale out of the technology platform for Aadhaar, the world’s largest identity program as a volunteer. He joined the Aadhaar project just when the first Aadhaar was issued, and was there when the billionth Aadhaar was generated. Vivek pioneered the use of AI at Aadhaar to improve the quality of services and detect identity fraud. Vivek continues to advise UIDAI on technology, in his volunteer role as Advisor(Technology).

Vivek had also served as volunteer CTO for Team Indus, India's entry to the Google Lunar X- Prize, which aimed to land a spacecraft on the surface of the moon.

Date: Tuesday, Sept 24, 2024
Time: 5:00 PM IST
Venue: AC01-LR-106
For details: ashoka-cdlds@ashoka.edu.in
or call: +91-9136857558

Viral V. Acharya is the C.V. Starr Professor of Economics in the Department of Finance at New York University Stern School of Business (NYU-Stern). He is a Research Associate of the National Bureau of Economic Research (NBER) in Corporate Finance, a Research Affiliate at the Center for Economic Policy Research (CEPR), and Research Associate of the European Corporate Governance Institute (ECGI). Viral was a Resident Scholar at the Federal Reserve Bank of New York (Sep 2022-Jan 2023) and a Deputy Governor at the Reserve Bank of India (RBI) during 23rd January 2017 to 23rd July 2019 in charge of Monetary Policy, Financial Markets, Financial Stability, and Research.

Viral's primary research interest is in theoretical and empirical analysis of systemic risk of the financial sector, its regulation, and its genesis in government- and policy-induced distortions, an inquiry that also examines the interaction of credit and liquidity risks, their agency-theoretic foundations, and their general equilibrium consequences.

Date: Tuesday, Sept 3, 2024
Time: 5:00 PM IST
Join online:
For details: ashoka-cdlds@ashoka.edu.in
or call: +91-9136857558