IndoML 2023 | IIT Bombay

Talk by Rajiv Khanna

Title: Interpretability in machine learning: Theory and Practice

Abstract: Continued and remarkable empirical successes of increasingly complicated machine learning models such as neural networks without a sound theoretical understanding of success and failure conditions can leave a practitioner blind-sided and vulnerable. As such, there has been an enhanced interest in recent times in research on building interpretable models as well as interpreting model predictions. In this talk, I will discuss some theoretical and practical aspects of interpretability in machine learning along both these directions through the lenses of feature attribution and example-based learning. In the first part of the talk, I will present novel theoretical results to bridge the gap in theory and practice for interpretable dimensionality reduction aka feature selection vis-à-vis the optimal svd-based matrix compression. I will discuss that the cost of interpretability accrued because of selecting features as opposed to principal components is not as high as was previously thought to be. In the second part of the talk, I will discuss the development of a framework and challenges associated with example-based machine learning for instance attribution to address ''Which training data points are responsible for making given test predictions'''. I will further discuss the importance of instance attribution in dissecting sharpness-based optimization.

Bio: Dr. Rajiv Khanna is currently Assistant Professor at the Department of Computer Science at Purdue University. Previously, he has held positions as a Visiting FacultyResearcher at Google Research, as a postdoc at the Department of Statistics at UC Berkeley, and a Research Fellow in the program of Foundations of Data Science at the Simons Institute for the Theory of Computing, also at UC Berkeley. His research is focused on elucidating mechanisms of success/failure conditions of machine learning through optimization, learning theory and interpretability. His work on beyond worst-case analysis on the Column Subset Selection won the best paper award at NeurIPS 2020. He completed his PhD in Electrical and Computer Engineering at UT Austin.

Speaker homepage

Talk by Pradeep Ravikumar

Title: Neuro-Causal Models

Abstract: First there was symbolic AI, with deterministic, logic and rule based systems, that were reliable but not that powerful. And then there was born statistical neural AI, with its distributions and uncertainties, that were powerful but not that reliable. What's next on the evolutionary ladder? Some argue that it will be a combination of the two, called neuro-symbolic AI. We discuss one instantiation of this we term neuro-causal models, that uses a synthesis of deep neural models and causal graphical models to automatically infer higher level symbolic information from lower level ''raw features''', while also allowing for rich relationships among the symbolic variables. Neuro-causal models retain the flexibility of modern deep neural network architectures while simultaneously capturing statistical semantics such as identifiability and causality, which are important to discuss ideal, target representations and their tradeoffs. We provide conditions under which this entire architecture can be recovered uniquely. We also discuss efficient algorithms and provide experiments illustrating the algorithms in practice.

Bio: Pradeep Ravikumar is a Professor in the Machine Learning Department, School of Computer Science at Carnegie Mellon University. He was previously an Associate Director at the Center for Big Data Analytics, at the University of Texas at Austin. His thesis has received honorable mentions in the ACM SIGKDD Dissertation award and the CMU School of Computer Science Distinguished Dissertation award. He is a Sloan Fellow, a Siebel Scholar, a recipient of the NSF CAREER Award. He is Associate Editor-in-Chief for IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), and incoming editor-in-chief of the Journal of Machine Learning Research.

Speaker homepage

Talk by Arya Mazumdar

Title: Iterative Hard Thresholding for Sparse Generalized Linear Models

Abstract: While linear models are well-understood in learning theory, non-linear models are still shrouded in some mystery, albeit being the powerhorse of machine learning for the past decade. The first step to understand non-linear learning problems is the study of generalized linear models that include linear regression, logistic regression, half-space learning, and two-layer neural networks as special cases. In this talk we discuss some algorithmic approaches for parameter learning in sparse generalized linear models (GLM). In particular, we focus on an iterative hard thresholding algorithm for sparse GLMs. We especially present convergence results for the binary iterative hard thresholding for the 1-bit compressed sensing problem, a very successful high-dimensional signal acquisition and recovery technique.

Bio: Arya Mazumdar is an Associate Professor of Data Science and Computer Science in UC San Diego. He is the Deputy Director and Associate Director for Research in the National AI Institute TILOS, and also the UCSD Site-Lead of NSF TRIPODS Institute EnCORE. Arya obtained his Ph.D. degree from University of Maryland, College Park specializing in information theory. Subsequently Arya was a postdoctoral scholar at Massachusetts Institute of Technology, an assistant professor in University of Minnesota, and an assistant followed by associate professor in University of Massachusetts Amherst. Arya is a recipient of a Distinguished Dissertation Award for his Ph.D. thesis, the NSF CAREER award, an EURASIP Best Paper Award, and the ISIT Jack K. Wolf Student Paper Award. He is also a Distinguished Lecturer of the IEEE Information Theory Society, 2023-24. He is currently serving as an Associate Editor for the IEEE Transactions on Information Theory and as an Area editor for Now Publishers Foundation and Trends in Communication and Information Theory. Arya's research interests include coding theory, information theory, statistical learning and optimization.

Speaker homepage

Talk by Krishna P. Gummadi

Title: Foundations of Foundational Models: Understanding the dynamics and mechanics of memorisation in large language models

Abstract: Understanding whether and to what extent large language models (LLMs) have memorised training data has important implications for the privacy of its training data and the reliability of its generated output. In this talk, I will focus on the more foundational question of how LLMs memorise training data. I will describe our work where we systematically trained LLMs of different sizes to memorise random token strings of different lengths and different entropies (i.e., sampled from different alphabet distributions) and study their ability to recall the strings. We observed many striking memorisation dynamics including (i) memorisation in phases with the alphabet distributions in the random strings being learnt before their relative positions in the string are memorised and (ii) memorisation in parts at the granularity of individual tokens, but not necessarily in the order in which they appear in the string. Next, we investigated memorisation mechanics by checking to what extent different parts of a token's prefix in the string are necessary and sufficient to recollect the token. I will show how we can leverage our insights to explain the dynamics of memorising strings and I will conclude by discussing the implications of our findings for quantifying memorisation.

Bio: Krishna Gummadi is a scientific director and head of the Networked Systems research group at the Max Planck Institute for Software Systems (MPI-SWS) in Germany. He also holds a professorship at the University of Saarland. He received his Ph.D. (2005) and B.Tech. (2000) degrees in Computer Science and Engineering from the University of Washington and the Indian Institute of Technology, Madras, respectively. Krishna's research interests are in the measurement, analysis, design, and evaluation of complex Internet-scale systems. His current projects focus on understanding and building social computing systems. Specifically, they tackle the challenges associated with (i) assessing the credibility of information shared by anonymous online crowds, (ii) understanding and controlling privacy risks for users sharing data on online forums, (iii) understanding, predicting and influencing human behaviors on social media sites (e.g., viral information diffusion), and (iv) enhancing fairness and transparency of machine (data-driven) decision making in social computing systems. Krishna's work on fair machine learning, online social networks and media, Internet access networks, and peer-to-peer systems has been widely cited and his papers have received numerous awards, including Test of Time Awards at ACM SIGCOMM and AAAI ICWSM, Casper Bowden Privacy Enhancing Technologies (PET) and CNIL-INRIA Privacy Runners-Up Awards, IW3C2 WWW Best Paper Honorable Mention, and Best Papers at NIPS ML & Law Symposium, ACM COSN, ACM/Usenix SOUPS, AAAI ICWSM, Usenix OSDI, ACM SIGCOMM IMC, ACM SIGCOMM CCR, and SPIE MMCN. He has also co-chaired AAAI's ICWSM 2016, IW3C2 WWW 2015, ACM COSN 2014, and ACM IMC 2013 conferences. He received an ERC Advanced Grant in 2017 to investigate "Foundations for Fair Social Computing".

Speaker homepage

Talk by Aditya Grover

Title: Foundation Models for Experimental Design

Abstract: Experimental design for optimizing black-box functions is a fundamental problem in many science and engineering fields. In this problem, sample efficiency is crucial due to the time, money, and safety costs of real-world design evaluations. Existing approaches either rely on active data collection or access to large, labeled datasets of past experiments, making them impractical in many real-world scenarios. In this work, we address the more challenging yet realistic setting of few-shot experimental design, where only a few labeled data points of input designs and their corresponding values are available. We introduce Experiment Pretrained Transformers (ExPT), a foundation model for few-shot experimental design that combines unsupervised learning and in-context pretraining. In ExPT, we only assume knowledge of a finite collection of unlabelled data points from the input domain and pretrain a transformer neural network to optimize diverse synthetic functions defined over this domain. Unsupervised pretraining allows ExPT to adapt to any design task at test time in an in-context fashion by conditioning on a few labeled data points from the target task and generating the candidate optima. We evaluate ExPT for few-shot experimental design in challenging domains and demonstrate its superior generality and performance compared to existing methods on a wide range of domains in Design Bench. Joint work with Tung Nguyen, Sudhanshu Agrawal, Siddarth Krishnamoorthy, and Satvik Mashkaria.

Bio: Aditya Grover is an assistant professor of computer science at UCLA. His goal is to develop efficient machine learning approaches that can interact and reason with limited supervision. He grounds this research in applications in climate and sustainability domains. Aditya's research works have been published at top venues including Nature, deployed in production at major technology companies, and covered in popular press venues such as Wall Street Journal and Washington Post. His research has been recognized with a best paper award at NeurIPS, prominent graduate research fellowships and faculty awards (Adobe, Google, Meta, Microsoft, Sony, Simons Institute), the ACM SIGKDD Doctoral Dissertation Award, and the AI Researcher of the Year Award by Samsung, and the Kavli Fellowship from the US National Academy of Sciences. Aditya received his postdoctoral training at UC Berkeley, PhD at Stanford, and bachelors at IIT Delhi, all in computer science.

Speaker homepage

Talk by Snigdha Chaturvedi

Title: Unsupervised methods for opinion summarization

Abstract: Online product reviews are important for both customers and businesses. However, the large number of reviews for a typical product can make it time-consuming to go through all of them. Opinion summarization aims to tackle this problem by creating a concise summary of the reviews. In this talk, I will discuss methods to perform extractive opinion summarization in an unsupervised manner. Our method uses dictionary learning to implicitly capture semantic information from the review and learns a latent representation of each sentence over abstract semantic units. Our method then leverages the representations to identify representative opinions among hundreds of reviews. I will also discuss methods to enhance this approach by automatically identifying the aspects described in the review sentences and then extracting sentences based on their aspects. We identify the underlying aspects of the review sentences using the roots of noun phrases and adjectives appearing in them. Our experiments show that such methods improve aspect coverage and produce more useful summaries.

Bio: Snigdha Chaturvedi is an Assistant Professor of Computer Science at the University of North Carolina, Chapel Hill. She specializes in Natural Language Processing, emphasizing narrative-like and socially aware understanding, summarization, and generation of language. Previously, she was an Assistant Professor at UC-Santa Cruz, and a postdoctoral fellow at UIUC and UPenn working with Dan Roth. She earned her Ph.D. in Computer Science from UMD in 2016, where she was advised by Hal Daume III. Her research has been supported by NSF, Amazon, and IBM.

Speaker homepage

Talk by Sachin Kumar

Title: NLP in the Time of Generative Models: Centering the Role of People

Abstract: Every language, like its speakers, is immensely diverse. Deep learning-based natural language processing models, even when trained on such heterogeneous data, are generally monolithic, encoding only the majority signal and averaging over all other variations. As a result, they consistently fail to support language use outside the ''standard'' and present challenges to the models' equitable access. In this talk, I discuss this issue in the context of generative models for text and describe how these shortcomings can be addressed by developing new adaptable and controllable training and inference algorithms. In the first part of the talk, I will discuss inference algorithms from pretrained autoregressive language models to control for stylistic and structural variations. I frame text generation as constrained optimization with gradient-based methods to generate text non-autoregressively, updating the entire output sequence iteratively. In the second part, I will introduce my recent work on diffusion-based text generation models that are non-autoregressive by design and have controllability baked in.

Bio: Sachin Kumar is a postdoctoral researcher at the Allen Institute for AI and the University of Washington, Seattle, and an incoming assistant professor at Ohio State University (starting in fall 2024). He received this Ph.D. from the Language Technologies Institute at Carnegie Mellon University, advised by Prof. Yulia Tsvetkov. Before starting his Ph.D., Sachin received his undergraduate degree in Computer Science from the Indian Institute of Technology Kharagpur. His research broadly revolves around Machine Learning and Natural Language Processing (NLP) with a particular focus on building NLP technologies that can account for language variation, that is those that can uniformly support diverse languages, domains, populations, and individuals.

Speaker homepage

Talk by Sunipa Dev

Title: Contextualizing AI with Cross Cultural Perspectives

Abstract: Development and evaluation of generative AI models rely heavily on semi-structured data annotated by humans. Both the data and human perspectives involved in the process, thus play a key role in what is taken as ground truth by models. Historically, this perspective has been Western-oriented which leads to a lack of representation of global contexts and identities in models as well as evaluation strategies, and the risk of disregarding marginalized groups that are most significantly affected by implicit harms. Accounting for cross-cultural differences in interacting with technology is an important step for building and evaluating AI holistically. We will talk through different strategies including community engaged approaches, along with survey experiments to capture a more diverse set of perspectives in data curation and benchmarking efforts. We will also zoom in on how socio-culturally aware AI research can fill in the gaps in fairness evaluations.

Bio: Sunipa Dev is a Senior Research Scientist at Google Research in its Responsible AI and Human Centered Technologies organization, working at the intersection of language, society, and technology. Previously, she was an NSF Computing Innovation Fellow at UCLA, before which she completed her PhD from the University of Utah. Her research strives to ground evaluations of generative AI, especially language technologies in real world experiences of people, and foster inclusion of diverse, cross-cultural, and marginalized perspectives into AI pipelines . She is also an organizer of Widening NLP at *CL conferences which argues the importance of diversity and inclusion in NLP for better technologies of the future.

Speaker homepage

Talk by Rishabh Iyer

Title: Subset Selection for Compute-Efficient Deep Learning: Orders of Magnitude Speedups with Subsets of the Training Data!

Abstract: In this Talk, we will cover some of our work (spanning the last 2+ years) on the use of subset selection for achieving compute-efficient training of deep learning and language model pretraining in several tasks and domains. In other words, instead of training on the entire subset, we will train on intelligently generated subsets of the dataset at every epoch. We will begin by providing a few basic paradigms of how we can achieve subset selection (a) GLISTER: a bi-level optimization based coreset selection, (b) GRAD-MATCH: a gradient matching based coreset selection, (c) CRAIG, which selects representative subsets using gradient embeddings, and (d) PRISM: a coreset selection scheme using parameterized submodular information measures. Next, we will provide theoretical guarantees for the subset selection, and we show how in fact, PRISM unifies and generalizes many of the above formulations (a - c) and also enables its use in scenarios where we have other desiderata beyond accuracy (e.g., robustness of imbalance, OOD data, distribution shift, or fairness). We will demonstrate the utility of our work in a number of settings including (1) supervised learning (Glister AAAI'21 and Grad-Match ICML'21, and MILO, Arxiv'23), (2) semi-supervised learning (Retrieve, NeurIPS'21), (3) hyper-parameter tuning and auto-ML including NAS (Automata, NeurIPS'22, AutoML WS'22, and MILO), (4) domain adaptation (Orient, NeurIPS'22), (5) speech recognition (EMNLP'22), and (6) INGENIOUS for compute-efficient LLM pre-training (EMNLP'23). In each case, we will show that we could achieve significant speedups (between 5x - 10x) with practically no loss in accuracy compared to training on the entire dataset, and efficiencies of 10x - 30x for hyper-parameter tuning and Auto-ML. This work not only reduces training time but also cost of compute (and correspondingly enables energy-efficient/green AI).

Bio: Prof. Rishabh Iyer is currently an Assistant Professor at the University of Texas, Dallas, where he leads the CARAML Lab. He is also a Visiting Assistant Professor at the Indian Institute of Technology, Bombay. He completed his Ph.D in 2015 from University of Washington, Seattle. He is excited in making ML more efficient (both computational and labeling efficiency), along with other desiderata like fairness, robustness, and so on. He has received the best paper award at Neural Information Processing Systems (NeurIPS/NIPS) in 2013, the International Conference of Machine Learning (ICML) in 2013, and an Honorable Mention at CODS-COMAD in 2021. He has also won the Adobe Data Science Research Award, a Microsoft Research Ph.D. Fellowship, a Facebook Ph.D. Fellowship, and the Yang Award for Outstanding Graduate Student from University of Washington.

Speaker homepage

Talk by Abhijit Sarkar

Title: Artificial intelligence in transportation: paving the way for the future of mobility

Abstract: In an era of rapid technological advancement, Artificial Intelligence (AI) has emerged as a driving force that is revolutionizing the landscape of transportation. The integration of AI technologies into various aspects of transportation research has sparked innovation, enhanced safety, and redefined efficiency in an industry that touches the lives of billions. We are learning to harness the power of big data, artificial intelligence, and high performance computing that can benefit all aspect of transportation. This presentation will delve into the profound impact of AI on surface transportation research. Starting from automated driving systems and traffic management to driver cognition and road safety, we will discuss some of the recent success stories including some of the recent work at Virginia Tech Transportation Institute. Together, we will embark on a journey to discuss the future of mobility. We will discuss the opportunities and challenges ahead, the power of innovation and collaboration, and the vision towards zero roadway fatality.

Bio: Dr. Abhijit Sarkar is a Senior Research Associate within the Division of Data and Analytics at the Virginia Tech Transportation Institute. At present, he leads the Computer Vision and Machine Learning team at VTTI. His research mainly focuses in the intersection of artificial intelligence and transportation research with an emphasis to computer vision. This includes safety and operations of automated driving systems, driver distraction and attention monitoring, intersection safety, unmanned aerial vehicles, multimodal sensor processing using naturalistic driving data, and AI application in driver health monitoring. He received his doctoral degree from the department of Electrical and Computer Engineering, Virginia Tech in 2017 with a focus in computer vision and remote health monitoring. Prior to that he received his Masters in Electrical Engineering from Indian Institute of Technology, Kharagpur, India and his Bachelor degree in Electrical Engineering from Jadavpur University, India. His research have been funded by multiple federal and state agencies including National Science Foundation (NSF), Federal Motor Carrier Safety Administrations (FMCSA), Federal Highway Administration (FHWA), National Highway Traffic Safety Administration (NHTSA), National Cooperative Highway Research Program (NCHRP), National Surface Transportation Safety Center for Excellence (NSTSCE), US Department of Transportation University Transportation Center (UTC), and multiple private companies.

Speaker homepage

Talk by Alekh Agarwal

Title: Improving generative models by playing games

Abstract: Improving generative models for natural language, code, images etc. typically involves a somewhat standard pipeline of pre-training on unsupervised data, subsequent supervised fine-tuning and a final phase of Reinforcement Learning from human preferences (RLHF) to improve the instruction-following properties of the models and steering them away from generating harmful content. This last step normally requires collecting human preferences among pairs of outputs on a diverse set of prompts according to some initial model, training a reward function to model these preferences, and subsequently using RL to optimize the reward function. The intermediate step of going from human preferences to rewards, however, is known to be brittle in situations when the preferences cannot be uniquely decomposed. In this work, we instead consider doing RL directly from either the human preference dataset, or a pairwise preference function trained on this data. We propose a novel game-theoretic formulation which captures good performance from the lens of this preference feedback. We show important structural properties of this game, which allow scalable solutions through self-play style strategies. We further analyze the convergence properties of our approach under various conditions on the preference feedback. We complement the theory by evaluating our methods on toy setups with challenging preference structures, preference feedback of varying difficulty in continuous control problems, and preference-based optimization of language models for text summarization. [Based on joint work with Gokul Swamy, Christoph Dann and Rahul Kidambi].

Bio: Alekh is a staff research scientist in the learning theory team at Google. Prior to that, he led the Reinforcement Learning group at Microsoft Research after getting his PhD from UC Berkeley. His research has been recognized with a neurips 2015 best paper award, icml 2022 best paper runner-up and an acm sigai industry impact award.

Speaker homepage

Talk by Mohit Bansal

Title: Multimodal Generative LLMs: Unification, Interpretability, Evaluation

Abstract: In this talk, I will present our journey of large-scale multimodal pretrained (generative) models across various modalities (text, images, videos, audio, layouts, etc.) and enhancing important aspects such as unification, efficiency, interpretability, and evaluation. We will start by discussing early cross-modal vision-and-language pretraining models (LXMERT). We will then look at early unified models (VL-T5) to combine several multimodal tasks (such as visual QA, referring expression comprehension, visual entailment, visual commonsense reasoning, captioning, and multimodal translation) by treating all tasks as text generation. We will also look at recent advanced unified models (with joint objectives and architecture, as well as newer unified modalities during encoding and decoding) such as textless video-audio transformers (TVLT), vision-text-layout transformers for universal document processing (UDOP), and composable any-to-any text-audio-image-video multimodal generation (CoDi). Second, we will discuss interpretable and controllable multimodal generation via LLM-based planning and programming, such as layout-controllable image generation via visual programming (VPGen), consistent multi-scene video generation via LLM-guided planning (VideoDirectorGPT), and open-domain, open-platform diagram generation (DiagrammerGPT). I will conclude with important evaluation aspects of multimodal generation models, based on fine-grained skill and social bias evaluation (DALL-Eval), as well as interpretable and explainable visual programs (VPEval).

Bio: Dr. Mohit Bansal is the John R. & Louise S. Parker Professor and the Director of the MURGe-Lab (UNC-NLP Group) in the Computer Science department at UNC Chapel Hill. He received his PhD from UC Berkeley in 2013 and his BTech from IIT Kanpur in 2008. His research expertise is in natural language processing and multimodal machine learning, with a particular focus on multimodal generative models, grounded and embodied semantics, language generation and Q&A/dialogue, and interpretable and generalizable deep learning. He is a recipient of IIT Kanpur Young Alumnus Award, DARPA Director's Fellowship, NSF CAREER Award, Google Focused Research Award, Microsoft Investigator Fellowship, Army Young Investigator Award (YIP), DARPA Young Faculty Award (YFA), and outstanding paper awards at ACL, CVPR, EACL, COLING, and CoNLL. He has been a keynote speaker for the AACL 2023 and INLG 2022 conferences. His service includes ACL Executive Committee, ACM Doctoral Dissertation Award Committee, CoNLL Program Co-Chair, ACL Americas Sponsorship Co-Chair, and Associate/Action Editor for TACL, CL, IEEE/ACM TASLP, and CSL journals.

Speaker homepage

Talk by Koustuv Saha

Title: Measuring Wellbeing in Situated Contexts with Social Media and Multimodal Sensing: Promises and Perils

Abstract: A core aspect of our social lives is often embedded in the communities we are situated in. Our shared experiences and social ties intertwine our situated context with our wellbeing. A better understanding of wellbeing can help devise timely support provisions. However, traditional forms of wellbeing measurements have limitations, motivating an increasing interest in supporting wellbeing through passive sensing technologies. Parallelly, social media platforms enable us to connect and express our personal and social lives with others. Given its ubiquity, social media can be considered a ''passive sensor'' to obtain naturalistic data, which can also be combined with various multimodal sensing to comprehensively measure wellbeing. However, wellbeing sensing technologies can lead to unintended outcomes and cause harms. Therefore, despite the potential, are we ready to deploy these wellbeing sensing technologies in the real world yet? In this talk, Koustuv Saha will present theory-driven computational and causal methods for leveraging social media in concert with complementary multisensor data to examine wellbeing, particularly in situated communities such as college campuses and workplaces. He will also interrogate the meaningfulness of the data and inferences and reflect on how these approaches can potentially be misinterpreted or misused without additional considerations. To bridge the gap between the theoretical promise and practical utility, he will present the importance of evaluating the needs, benefits, and harms of wellbeing sensing technologies in practice. This talk will propel the vision toward questioning the underlying assumptions and in responsible design and deployment of wellbeing sensing technologies (if at all) for situated communities and the future of work.

Bio: Koustuv Saha is an Assistant Professor of Computer Science at the University of Illinois at Urbana-Champaign (UIUC). He was previously a Senior Researcher at Microsoft Research, Montreal, in the Fairness, Accountability, Transparency, and Ethics in AI (FATE) group, and he completed his Ph.D. in Computer Science from Georgia Tech in 2021. His research interest is in social computing, computational social science, human-centered machine learning, and FATE. He adopts machine learning, natural language, and causal inference analysis to examine human behavior and wellbeing using different forms of digital data, including social media and multimodal sensing data. His work questions the underlying assumptions of data-driven inferences and the possible harms such inferences might lead to. His research is situated in an interdisciplinary and human-centered context and bears implications for various stakeholders. His work has been published at various venues, including CHI, CSCW, ICWSM, IMWUT (UbiComp), Scientific Reports, JMIR, FAccT, among others. He is a recipient of the 2021 Outstanding Doctoral Dissertation Award from the College of Computing at Georgia Tech, Foley Scholarship Award from the GVU Center, Snap Research Fellowship, and a finalist of the Symantec Graduate Fellowship. His research has won the Outstanding Study Design Award at ICWSM, and has been covered by several media outlets, including the New York Times, Vox, CBC Radio, NBC, 11Alive, the Hill, and the Commonwealth Times. During his Ph.D., he did research internships at Snap Research, Microsoft Research, Max Planck Institute, and Fred Hutch Cancer Research. Earlier, he completed his B.Tech (Hons.) in Computer Science and Engineering from the Indian Institute of Technology (IIT) Kharagpur. He was awarded the NTSE Scholarship by the Govt. of India, and he has over six years of industry research experience.

Speaker homepage

Talk by Vishnu Vardhan

Title: Applications of AI in Metabolic Health

Abstract : The integration of Artificial Intelligence (AI) in the realm of metabolic health has emerged as a transformative force, revolutionizing the understanding, diagnosis, and management of metabolic disorders. Leveraging machine learning, AI plays a crucial role in predictive analytics, enabling early detection and personalized risk assessment for metabolic disorders. Deep learning models contribute to precision medicine by tailoring treatment plans based on individual genetic profiles, optimizing efficacy, and minimizing side effects. This talk explores the multifaceted applications of AI in addressing the complexities of metabolic health, ranging from personalized diagnostics to innovative therapeutic interventions. In summary, this talk highlights how AI is transforming the landscape of metabolic health, offering personalized insights, early interventions, and innovative approaches to redefine healthcare practices.

Bio: Dr Vishnu Vardhan is an Orthopedic and Ballistic Trauma surgeon transformed into digital health evangelist & HealthTech entrepreneur. He is the Founder & CEO of Vizzhy Inc, a pioneer company which is using Multiomics approach to manage metabolic health. He is the Co- founder of CARE365 and working with US Medicare in solving their complex care management problems using IoTs, AI and Virtual patient management He holds double masters from Massachusetts institute of technology (MIT). He has worked on Artificial intelligence, IoTs, healthcare analytics, Big Data and Blockchain technology and consulted in Mass General Hospital, Harvard Medical School, Tufts University and BCG.

Speaker homepage

Talk by Purnamrita Sarkar

Title: Keep or toss? A nonparametric score to evaluate solutions for noisy ICA

Abstract: Independent Component Analysis (ICA) was introduced in the 1980's as a model for Blind Source Separation (BSS) by Comon et al, which refers to the process of recovering the sources underlying a mixture of signals, with little knowledge about the source signals or the mixing process. It has become a powerful alternative to the Gaussian model leading to PCA in factor analysis. The key idea in ICA is the development of a contrast function which is minimized for Gaussian random vectors and leads to an objective function that is maximized when one learns the directions of the independent components. The major algorithms FastICA (Hyvarinen, 1999) and JADE (Cardoso, 1993) do this using functions based on kurtosis. These algorithms can give severely biased results if the independent component distributions are heavy-tailed or have no higher moments (Anderson et al., 2017; Chen and Bickel, 2005; Hyvarinen, 1997) or the kurtosis of the components is small in magnitude. We consider the noisy ICA, where one has a linear combination of independent non-Gaussian random variables perturbed by independent additive Gaussian noise. While there are many works with different contrast functions and different sophisticated algorithms for estimation, we develop a diagnostic to adaptively pick the right algorithm to uncover the directions from a set of available algorithms. In addition, we propose some new contrast functions and algorithms that enjoy the same fast computability as existing algorithms like FASTICA and JADE but work in domains where the former may fail. While these also may have weaknesses, our proposed diagnostic, as shown by our simulations, can remedy them. This is joint work with Syamantak Kumar and Peter Bickel.

Bio: Purnamrita Sarkar is an associate professor at the Department of Statistics and Data Science at the University of Texas at Austin. Prior to joining UT Austin, she received her Ph.D. in Machine Learning from Carnegie Mellon University and was a postdoctoral scholar at the Departments of Statistics and Electrical Engineering and Computer Science at UC Berkeley. Dr. Sarkar is interested in foundational problems arising in networks, resampling, streaming algorithms, and general unsupervised learning. Her work typically blends tools from theoretical statistics, optimization, machine learning and random matrix theory. At UT Austin, she co-leads the NSF Tripods Phase 2 Institute and is affiliated with the NSF AI Institute for Foundations of Machine Learning.

Speaker homepage