Tutorials

Breaking the Memory Wall: Towards Efficient Long Context and Chain-of-Thought Reasoning

Souvik Kundu (Intel Labs USA)
9:30 AM -12 PM | Lecture Room 101

Since their inception, transformer based foundation models like LLMs have become increasingly popular, particularly after the "ChatGPT" moment. These models have accelerated the growth and adaptation of the foundational AI models in various use cases including in automated chatbot, quantitative trading, and AI genomics. However, with the rapid adaptation of these billion-parameter scale models, the demand for memory, compute, and power has skyrocketed. This raises a fundamental question on the sustainability and inclusivity of such growth. There have been continuous research efforts to mitigate this pressing problem and make AI more affordable. This tutorial will focus on providing a general landscape on the state of efficient foundation models in the spectrum of sustainable AI systems development. Specifically, the tutorial will have three main components. First, it will start by covering the key inference-time optimizations for LLMs to mitigate their computation-memory bottlenecks while improving the tokens/$. Secondly, the tutorial will cover the advancements of LLMs in the space of fine-tuning and personalization. Finally, the tutorial will conclude upon discussion on the advanced model optimizations that are essential to push the LLMs to process longer contexts of data and unleash efficient reasoning capabilities.

Data and Quantum Machine Learning

M. S. Santhanam (IISER Pune), Nisarg Vyas (IISER Pune)
9:30 AM -12 PM | Lecture Room 103

Quantum machine learning is an emerging area that seeks to exploit the quantum effects – superposition and entanglement – for speeding up classical machine learning algorithms. In the last one decade, a host of quantum algorithms have been shown to display significant speedups – ranging from polynomial to exponential speedup – and improved accuracy over their corresponding classical algorithms and better accuracy. Thus, QML promises to deliver better overall performance for solving a variety of machine learning tasks. For instance, quantum version of support vector machine (qSVM) is shown to be exponentially faster compared to the classical SVM. The first part of this tutorial will provide an overview of these theoretical developments in quantum machine learning. The second part will focus on how classical data must be encoded into quantum states to realize the full potential of these algorithms. In the third part, through a series of examples, we will discuss the gaps between the theoretical results and practical implementations. Finally, using python based frameworks, we will demonstrate data encoding and working of one or two algorithms using selected data sources.

This tutorial is an overview and does not require the audience to have a background in quantum physics. It is suitable for anyone with an interest in novel approaches to machine learning problems.

Towards Fair and Trustworthy AI: Foundations, Challenges, and Future Directions

Puspita Majumdar (AI Garage, Mastercard), Balraj Prajesh (AI Garage, Mastercard), Nitendra Rajput (AI Garage, Mastercard), and Ankur Arora (AI Garage, Mastercard)
9:30 AM -12:00 PM | Lecture Room 105

Machine learning algorithms are becoming an increasingly integral part of critical decision-making processes across diverse sectors, including hiring, finance, university admissions, and criminal justice. While these systems offer unprecedented efficiency, they carry a significant vulnerability: they may inherit, perpetuate, and even amplify subtle societal biases embedded in the data on which they are trained. These biases often originate from historical inequities, imbalanced class distributions, or spurious correlations with sensitive attributes such as race, gender, or age. The consequences are profound and well-documented, ranging from discriminatory loan denials and inequitable job screenings to flawed recidivism predictions that disproportionately harm marginalized communities. Beyond raising serious ethical concerns, such biases also expose organizations to substantial legal and regulatory risks. The challenge has become even more acute with the widespread adoption of large language models (LLMs) and their derivative systems, including chatbots, retrieval-augmented generation frameworks, and autonomous agents, which reproduce these biases directly in natural language, thereby shaping user interactions and information consumption in particularly visible and impactful ways. In this context, this tutorial provides a comprehensive introduction to fairness in machine learning and AI systems. It highlights the need to think about fairness in a broader way, rather than focusing only on accuracy. We will look at how bias can appear at different stages of the model development process. The tutorial will also cover frameworks that help evaluate fairness, discuss important fairness metrics, and the trade-off between fairness and accuracy.

Seeking Deep, Building Big: A Founder’s Guide to Starting Up

Monish Darda (Icertis), Harshad Oak (Icertis)
1:30 PM - 4:00 PM | Lecture Room 101

This session explores what it takes to move from an idea to a company that scales globally. Drawing on firsthand lessons from building multiple startups, including Icertis, a global leader in AI-powered contract lifecycle management, this session includes candid stories, hard lessons, and practical frameworks for product market fit, funding choices, team building, and scaling operations.

The session is highly interactive, combining storytelling, hands-on exercises, and open discussion. Attendees will work through founder dilemmas, stress-test their ideas, and align lessons with their personal goals. By the end, they will have practical tools, a sharper perspective, and the confidence to seek deep and build big.

Foundations of Generative AI

Indrajit Bhattacharya (KnowDis AI)
1:30 PM - 4:00 PM | Lecture Room 103

This 'tutorial' is more of a guided tour exploring the foundations of generative AI from many different points of view, with the goal of dispelling prevailing 'gen AI'-related myths and misunderstandings. We will begin by defining generative AI, understanding what distinguishes it from its 'non-generative' counterpart (yes, that exists!) and appreciating how generative AI approaches the essence of human and superhuman agency. We will then define the core tasks in generative AI, namely, sampling, scoring, inference and learning, in addition to the meta-task of modeling. We will analyze the key challenges involved in each of these, and the intrinsic trade-offs involved in addressing these simultaneously. We will then turn our gaze backwards in time and look far beyond 2012 (yes, AI did exist back then!), beyond even the 1950s, to appreciate the debt that contemporary researchers and developers owe to giants from different cultures, continents, disciplines and eras for the tools and techniques used to address these core tasks. Next, we will review the building blocks of generative AI to understand how cutting-edge generative AI models of today borrow from and build upon traditional generative models (no, I don't mean GPT-1!). Specifically, we will review generative models for classification and regression, IID versus structured models, conditional independence and the graphical model view, latent variables (mixture models versus factor analysis), and the troublesome partition function. Then, and only then (yes, we are taking the long route!), we will review a few specific categories of generative models --- autoregressive models, variational autoencoders, generative adversarial networks, normalizing flows, and diffusion models --- and analyze how these arrive at different compromises between the competing goals of generative modeling. We will end by thinking through the significant risks posed by generative AI alongside transformative benefits, and discuss the deep ethical considerations involved in building, owning and using generative AI models.

AI, Agents and Software Engineering - What’s Happening and Where Are We Heading?

Sridhar Chimalakonda (IIT Tirupati) and Atul Kumar (IBM Research)
1:30 PM - 4:00 PM | Lecture Room 105

The intersection of Artificial Intelligence (AI) and Software Engineering (SE) has emerged as a significant research area with substantial practical implications for industry as well. There is immense work that addresses software engineering challenges (e.g. requirements generation, code generation, code review, code completion, bug fixing, bug localization), and also a lot of work that aims to address challenges in AI systems through software engineering. Software engineers (also AI engineers, data scientists) today apply AI tools (e.g., Claude, CoPilot) to create, analyze, evaluate, and process a range of diverse software artifacts, including source code, requirements documents, architectural specifications, test cases, and maintenance logs across the entire software development lifecycle. This tutorial aims to provide an overview of the landscape of AI-driven approaches to software engineering, with a focus on autonomous software development agents and research directions for future developments. The tutorial covers three primary areas: (1) Current State-of-the-art: How AI models address a range of software engineering tasks ranging from requirements analysis to testing (2) Agent-Based Systems: landscape of AI agents capable of performing end-to-end software engineering tasks across multiple development phases (3) Future Research Directions: Can we trust and rely on AI-generated software? How to address explainability and security challenges? How to deal with human-AI collaboration in the context of AI+SE?