Carnegie Mellon University
Tuesday, October 5, 2021, 3:00 PM EDT (1:00 PM MDT)
A video of this talk is available to NIST staff in the Math channel on NISTube, which is accessible from the NIST internal home page.
Abstract: Motivated by networks of neuroscience and AI, this talk delves into the question of how to examine computational systems, and influence them to attain desirable outcomes. A central and widely studied problem in neuroscience is to understand the representation and flow of information in the healthy and diseased brain. The first part of my talk will draw from existing neuroscience literature and simple examples to arrive at the first formal definition of information flow in the brain, developing what we call the “M-Information Flow” framework. With this, I will show that it is possible to verifiably track information flow about a given message, and, further, obtain finer grained information than existing tools. To that end, we leverage recent developments in information theory on "Partial Information Decomposition" (PID), which provides new tools for defining and estimating unique, redundant and synergistic information. We will also see how synergy can arise in unexpected ways in the brain, through a case study of the so-called “grid cells”. Separately, through examples, I will also illustrate how the M-Information Flow framework can be used to arrive at precise interventions on a network to control the influence of different messages on the output, and use this result as a motivation to discuss the formalization and the undecidability of the general problem of finding minimal interventions.
Towards translation to the human brain, I will discuss fundamental limits and novel problems in noninvasive sensing and stimulation of the brain, including a new problem of localizing “silences” in the brain (with applications to predicting worsening brain injuries using the first noninvasive detection of waves called “cortical spreading depolarizations”), and designing the first neural sensing systems that work with Black hair types. I will also illustrate how information and control-theoretic approaches are beginning to influence neuromodulation, and discuss novel machine-learning techniques to find waveforms that perform localized and cell-type selective neurostimulation.
Finally, I will briefly discuss how this information flow perspective and PID also offers novel ways of thinking about explainability problems of critical importance in fairness, accountability, trust, and explainability of machine learning. One problem of my focus is of allowing exemptions in biases (e.g. biases that are explainable as “business necessities” are exempt in the US hiring laws). This motivates an important question: how can we quantify the exempt and non-exempt biases in a machine-learnt model? I will overview our measure that satisfies certain desirable properties, and synthesizes ideas from PID and causality. I will discuss how this measure can be applied in practical setups to both make these decisions fairer and consistent with the law, and how we are using it to understand bias in our graduate admissions.
The talk largely is based on joint works with my students, Sanghamitra Dutta, Praveen Venkatesh, Alireza Chamanzar, Chaitanya Goswami, and Arnelle Etienne.
Bio: Pulkit (Ph.D. UC Berkeley'10, B.Tech, M.Tech IIT Kanpur) is the Angel Jordan Associate Professor at CMU. His main contributions to science are towards developing and experimentally validating new science of information for energy-efficient communication, distributed control, neural inference, sensing, and stimulation. To apply these ideas to a variety of problems including ethical AI and novel biomedical systems, his lab works extensively with data scientists, system and device engineers, neuroscientists, and clinicians. Specifically, work of his foundations of AI and neuroengineering lab is focused on a) fair and explainable AI at algorithm, theory, and hardware level; b) tools (theoretical, computational, and hardware) for understanding the healthy brain, and understanding, diagnosing, and treating disorders such as epilepsy, stroke, and traumatic brain injuries. Pulkit received the 2010 best student paper award at IEEE Conference on Decision and Control; the 2011 Eli Jury Dissertation Award from UC Berkeley; the 2012 Leonard G. Abraham best journal paper award (IEEE ComSoc); a 2014 NSF CAREER award; a 2015 Google Research Award; a 2018 inaugural award from the Chuck Noll Foundation for Brain Injury Research; the 2018 Spira Excellence in Teaching Award (CMU), and the 2019 best tutorial paper award (IEEE ComSoc). He's the CEO of Precision Neuroscopics, Inc., a startup translating his work on accessible, high-resolution neural sensing solutions to the real world. He's learning how to play the sax and enjoys his free time with his wife and children.
Note: This talk will be recorded to provide access to NIST staff and associates who could not be present to the time of the seminar. The recording will be made available in the Math channel on NISTube, which is accessible only on the NIST internal network. This recording could be released to the public through a Freedom of Information Act (FOIA) request. Do not discuss or visually present any sensitive (CUI/PII/BII) material. Ensure that no inappropriate material or any minors are contained within the background of any recording. (To facilitate this, we request that cameras of attendees are muted except when asking questions.)
Host: Tony Kearsley