Applications of Computational Math In Industry Spring 2026 (CME 500)

Guest Speaker:
Alex Infanger, ICME Alum, Postdoctoral Researcher

Abstract:

Because human preferences are too complex to codify, AIs operate with misspecified objectives. Optimizing such objectives often produces undesirable outcomes; this phenomenon is known as reward hacking. Such outcomes are not necessarily catastrophic. Indeed, most examples of reward hacking in previous literature are benign. And typically, objectives can be modified to resolve the issue. We study the prospect of catastrophic outcomes induced by AIs operating in complex environments. We argue that, when capabilities are sufficiently advanced, pursuing a fixed consequentialist objective tends to result in catastrophic outcomes. We formalize this by establishing conditions that provably lead to such outcomes. Under these conditions, simple or random behavior is safe. Catastrophic risk arises due to extraordinary competence rather than incompetence. 

This talk is based on recent work https://arxiv.org/abs/2603.15017

Guest Speaker: 
Rajesh Parekh, Vice President of AI/ML

Abstract:
Intuit is building an AI-driven expert platform that leverages Artificial Intelligence and Human Intelligence (AI+HI) to unlock financial opportunities for over 100 million customers world-wide. This talk focuses on the fundamental advancements in AI technology to power “done for you” experiences in the FinTech space. Specifically, I will present key research ideas and insights that enabled large-scale document understanding and improved the performance ceiling of agentic systems.

Guest Speakers:
Jian Jiao, Partner Applied Science Manager
Vishnu Navda, Partner Engineering Manager 

Abstract:
This presentation examines the evolution of monetization systems in industry, from keyword search and ad ranking to neural retrieval, recommendation, commerce, and emerging agentic experiences. We will briefly walk through the core ideas behind these systems, and then focus on modern monetization with an emphasis on advertising and commerce.

Guest Speaker:
Matt Hancher, Director of Engineering for the Geo for the Environment

Abstract:
At Google we've been making geospatial data accessible and useful since the dawn of Google Earth and Maps twenty years ago. Now deep learning and generative AI are transforming the geosciences again, reshaping how we approach everything from mapping and monitoring agriculture and forests to forecasting the weather and responding to natural disasters. This talk will explore how these new AI techniques — including the breakthrough model known as AlphaEarth Foundations — are applied to a range of sustainability challenges such as monitoring deforestation-free supply chains and establishing and monitoring protected areas. We will explore how AI is unlocking questions we've never been able to answer before, and how it's radically lowering the barrier to entry so everyone can make better decisions about our planet.

Guest Speaker:
Hari Sowrirajan, Research Engineer

Abstract:
In this talk, Hari will present Aristotle, the automated theorem-proving system that achieved Gold-medal-level performance at last year's International Mathematical Olympiad (IMO). He will dive into the mechanics of formal reasoning with Lean and explore how it integrates with proof search, reinforcement learning, and test-time scaling. Beyond the technical architecture, he will examine how these systems are already accelerating mathematical research and their potential to redefine collaboration—both between humans and machines.

Guest Speaker:
Iraj Hosseini, Distinguished Scientist and Director/Pharmacology Team Leader

Abstract: 
Digital twins, virtual representations of individual clinical patients, are emerging as a powerful approach in systems pharmacology and drug development. By integrating genomics, physiological measurements, laboratory data, and other clinical datasets, digital twins can capture patient-specific biological and pharmacological characteristics. For Quantitative Systems Pharmacology (QSP) applications, digital twins help address a key challenge in clinical trial simulation: accurately representing patient variability, particularly for novel therapies or new patient populations where parameter distributions are uncertain. By generating individualized model parameterizations, digital twins enable the simulation of treatment responses, evaluation of clinical outcomes with limited data, and exploration of numerous “what-if” scenarios. In this talk, we’ll discuss the applications of digital twins, which have been successfully used in early clinical development, including Phase 1 studies of T-cell–engaging bispecific antibodies to characterize dose–response relationships and identify predictive biomarkers, as well as in TCR-engineered cell therapies to model T-cell kinetics and predict the impact of product composition on patients’ T-cell responses. Together, digital twins and QSP modeling support more predictive clinical trial design and accelerate the development of patient-centric therapeutic strategies.
 

Guest Speaker:
Frédéric Poitevin, Staff Scientist

Students are expected to attend the ICME Research Symposium in lieu of a class lecture.

The ICME Research Symposium takes an in-depth look at current research and future directions in areas like computational mathematics, data science, machine learning, and scientific computing, with perspectives from industry professionals, faculty, and student scholars.
View the latest agenda.

Guest Speaker:
Shalin Mehta, Senior Manager AI/ML

Guest Speaker:
Danielle Maddix Robinson. Senior Applied Scientist

Abstract:
In this talk, we discuss the bitter lesson in designing time series foundation models (TSFMs). First, we introduce Chronos and Chronos-Bolt models and how they differ in their design choices. Importantly, we use these models to more broadly represent general design choice differences in TSFMs, e.g., patch size, continuous vs. quantization embedding, and regression vs. classification loss function. We then show that while Chronos-Bolt, which has more natural time series inductive biases, e.g., continuous embedding and quantile loss function, performs better on classical time series benchmarks, Chronos performs better on chaotic systems. We then identify biases induced by these design choices, e.g., temporal, geometry and regression-to-the-mean biases to explain what is causing these different behaviors and the pros/cons of each design choice. Lastly, we conclude with forward looking view on TSFMs and the newly-released multivariate Chronos-2 TSFM.