Jonathan Rosenblatt, PhD- Senior Lecturer (Asst. Prof), Department of Industrial Engineering and Management, Ben Gurion University, Negev, Israel
When: August 7th, 9:45-11:00 am
Title: On the low power of predictive accuracy for signal detection
Abstract: The estimated accuracy of a supervised classifier is a random quantity with variability. A common practice in supervised machine learning, is thus to test if the estimated accuracy is significantly better than chance level. This method of signal detection is particularly popular in neuroimaging and genetics. We provide evidence that using a classifier's accuracy as a test statistic can be an underpowered strategy for finding differences between populations, compared to a bona-fide statistical test. It is also computationally more demanding. We compare test statistics that are based on classification accuracy, to others based on multivariate test statistics. We find the probability of detecting differences between two distributions is lower for accuracy-based statistics. We examine several candidate causes for the low power of accuracy tests. These causes include: the discrete nature of the accuracy test statistic, the type of signal accuracy tests are designed to detect, their inefficient use of the data, and their regularization. When the purpose of the analysis is not signal detection, but rather, the evaluation of a particular classifier, we suggest several improvements to increase power. In particular, to replace V-fold cross validation with the Leave-One-Out Bootstrap.
Biography: Jonathan D. Rosenblatt Is a Senior Lecturer (Assistant Professor) in the Dept. of Industrial Engineering and Management, Ben Gurion University of the Negev, Israel. He is a statistician, working on wide range of topics and applications including distributed algorithms for machine-learning, statistical methods for medical imaging, statistical theory, high-dimensional process control, and more.
Max Welling, PhD- Professor of Computer Science, Institute of Informatics, University of Amsterdam, Netherlands
When: July 24, 9:45-11:00 am
Title: Graph Nets: The Next Generation
Abstract: In this talk I will introduce the next generation of graph neural networks known as Natural Graph Networks and Mesh CNNs. GNNs have the property that they are invariant to permutations of the nodes in the graph. This turns out to be unnecessarily limited. In this work we develop new models that are more flexible in the sense that they do not have isotropic kernels but at the same time remain highly scalable. The Mesh-CNNs are developed to run messages over a graph (mesh) that represents a discretization of a (curved) surface. The Natural Graph Networks are designed to be more flexible convolutions on general graphs. Joint with Pim de Haan, Maurice Weiler, and Taco Cohen.
Biography: Prof. Dr. Max Welling is a research chair in Machine Learning at the University of Amsterdam and a VP Technologies at Qualcomm. He has a secondary appointment as a fellow at the Canadian Institute for Advanced Research (CIFAR). Max Welling has served as associate editor in chief of IEEE TPAMI from 2011-2015. He serves on the board of the Neurips foundation since 2015 and has been program chair and general chair of Neurips in 2013 and 2014 respectively. He was also program chair of AISTATS in 2009 and ECCV in 2016 and general chair of MIDL 2018. He is a founding board member of ELLIS. Max Welling is recipient of the ECCV Koenderink Prize in 2010. He directs the Amsterdam Machine Learning Lab (AMLAB), and co-directs the Qualcomm-UvA deep learning lab (QUVA) and the Bosch-UvA Deep Learning lab (DELTA). He has over 300 publications in machine learning and an h-index of 66.
Alexander Lavin- Chief Scientific Officer, Augustus Intelligence, NY, NY, Founder, Latent Sciences, Cambridge, MA
When: July 17th, 9:45-11:00 am
Title: Machine learning research to production in medicine
Abstract: Taking machine learning (ML) models and algorithms from R&D to production is often non-trivial, and exponentially so in medical applications: real-world patient data is typically ill-prepared for ML, deployment settings vary in often subtle ways that affect data distributions and thus model performance, model interpretability and explainability at several levels of abstraction are needed for usability and trust, principled uncertainty reasoning is critical for confidence in practice, and more. The tasks and datasets in ML research rarely reflect the real-world objectives and constraints. In this talk I discuss the misalignment issue with ML research and applications in medicine, and specifically prescribe ways to advance medical ML R&D to real-world deployment. I elucidate this with several examples: developing a state-of-the-art neurodegenerative prediction algorithm towards a personalized medicine application, and a novel unsupervised computer vision method for use in neuropathology.
Biography: Alexander Lavin is an AI researcher and software engineer, specializing in Bayesian machine learning and probabilistic computation. Lavin is Chief Scientific Officer at stealth Augustus Intelligence, building state-of-art "augmented intelligence" for massive real-world challenges. Lavin is also founder of Latent Sciences, a startup commercializing his patented AI platform for predictive disease modeling; the flagship application is presymptomatic prediction of neurodegeneration. Before Augustus and Latent, he was a Senior Research Engineer at both Vicarious and Numenta, building artificial general intelligence for robotics, and developing biologically-derived AI & ML algorithms, respectively. He was previously a spacecraft engineer, and now is an AI Advisor for NASA FDL. Lavin was a Forbes 30 Under 30 honoree in Science, advises several deep tech startups (from next-gen computation to medical devices), and has published in top journals and conferences across AI/ML and neuroscience. In his free time, Alexander enjoys running, yoga, live music, and reading sci-fi and theoretical physics books.
Olaf Sporns, PhD- Professor and Chair, Department of Psychological and Brain Sciences, Indiana University, Bloomington IN
When: July 10th, 9:45 - 11:00 am
Title: Connectivity and Dynamics of Complex Brain Networks
Abstract: Networks (connectivity) and dynamics are two key pillars of network neuroscience – an emerging field dedicated to understanding structure and function of neural systems across scales, from neurons to circuits to the whole brain. In this presentation I will review current themes and future directions, including structure/function relationships, use of computational models to map information flow and communication dynamics, and a novel edge-centric approach to functional connectivity. I will argue that network neuroscience represents a promising theoretical framework for understanding the complex structure and functioning of nervous systems.
Biography: After receiving an undergraduate degree in biochemistry, Olaf Sporns earned a PhD in Neuroscience at Rockefeller University and then conducted postdoctoral work at The Neurosciences Institute in New York and San Diego. Currently he is the Robert H. Shaffer Chair, a Distinguished Professor, and a Provost Professor in the Department of Psychological and Brain Sciences at Indiana University in Bloomington. He is co-director of the Indiana University Network Science Institute and holds adjunct appointments in the School of Informatics and Computing and the School of Medicine. His main research area is theoretical and computational neuroscience, with a focus on complex brain networks. In addition to over 200 peer-reviewed publications he is the author of two books, “Networks of the Brain” and “Discovering the Human Connectome”. He is the Founding Editor of “Network Neuroscience”, a journal published by MIT Press. Sporns was awarded a John Simon Guggenheim Memorial Fellowship in 2011, was elected Fellow of the American Association for the Advancement of Science in 2013, and received the Patrick Suppes Prize in Psychology/Neuroscience, awarded by the American Philosophical Society in 2017.
Omer T Inan, PhD- Associate Professor, Electrical and Computer Engineering, Georgia Tech, Atlanta, GA
When: June 12, 10:15-11:30 am
Title: Non-Invasive Physiological Sensing and Modulation for Human Health and Performance
Abstract: Recent advances in digital health technologies are enabling biomedical researchers to reframe health optimization and disease treatment in a patient-specific, personalized manner. Rather than a one-size-fits-all paradigm, the charge is for a particular profile to be fit to each patient, and for disease treatment (or wellness) strategies to then be tailored accordingly—perhaps even with fully closed-loop systems based on neuromodulation. Non-invasive physiological sensing and modulation can play an important role in this effort by augmenting existing research in ‑omics and medical imaging towards better developing such personalized models and phenotypic assays for patients, and in continuously adjusting such models to optimize therapies in real-time to meet patients’ changing needs. While in many instances the focus of such efforts is on disease treatment, optimizing performance for healthy individuals is also a compelling need. This talk will focus on my group’s research on non-invasive sensing of the sounds and vibrations of the body, with application to musculoskeletal and cardiovascular monitoring applications. In the first half of the talk, I will discuss our studies that are elucidating mechanisms behind the sounds of the knees, and particularly the characteristics of such sounds that change with acute injuries and arthritis. We use miniature microelectromechanical systems (MEMS) air-based and piezoelectric contact microphones to capture joint sounds emitted during movement, then apply data analytics techniques to both visualize and quantify differences between healthy and affected knees. In the second half of the talk, I will describe our work studying the vibrations of the body in response to the heartbeat using wearable MEMS accelerometers, and how this sensing fits within a non-invasive neuromodulation ecosystem for treating post-traumatic stress disorder. Our group has extensively studied the timings of such vibrations in relation to the electrophysiology of the heart, and how such timings change for patients with cardiovascular diseases during treatment. Ultimately, we envision that these technologies can enable personalized titration of care and optimization of performance to reduce injuries and rehabilitation time for athletes and soldiers, improve the quality of life for patients with heart disease, and reduce overall healthcare costs.
Biography: Omer Inan is an Associate Professor of Electrical and Computer Engineering and Adjunct Associate Professor of Biomedical Engineering at Georgia Tech. He received his BS, MS, and PhD in Electrical Engineering from Stanford in 2004, 2005, and 2009, respectively. From 2009-2013, he was the Chief Engineer at Countryman Associates, Inc., a professional audio manufacturer of miniature microphones and high-end audio products for Broadway theaters, theme parks, and broadcast networks. He has received several major awards for his research including the NSF CAREER award, the ONR Young Investigator award, and the IEEE Sensors Council Early Career award. While at Stanford as an undergraduate, he was the school record holder and a three-time NCAA All-American in the discus throw.
Rajesh Ranganath, PhD- Assistant Professor, Courant Institute of Mathematical Sciences and the Center for Data Science, New York University, New York, NY
When: May 29, 9:45-11 am
Title: Checking AI via Testing and its Application to COVID-19
Abstract: AI powers predictive models of medical data by uncovering subtle relationships between the input to the model and its output. However, the power of AI models means they can pick up on spurious relationships in data. Methods to surface these relationships can make AI models more robust. In the first part of this talk, I will show how powerful probabilistic models can be used to build hypothesis tests that identify important relationships in data. Along the way, I will discuss techniques for highlighting important pieces of the input for a particular observation. In the second part of the talk, I will discuss how these approaches have been used to support the development of an adverse event model for hospitalized COVID-19 patients and a visualization of this model for clinicians at the bedside.
Biography: Rajesh Ranganath is an assistant professor at NYU's Courant Institute of Mathematical Sciences and the Center for Data Science. He is also affiliate faculty at the Department of Population Health at NYUMC. His research focuses on approximate inference, causal inference, Bayesian nonparametrics, and machine learning for healthcare. Rajesh completed his PhD at Princeton and BS and MS from Stanford University. Rajesh has won several awards and fellowships including the NDSEG graduate fellowship, the Porter Ogden Jacobus Fellowship, given to the top four doctoral students at Princeton University, and the Savage Award in Theory and Methods.
Julia Schnabel, PhD- Professor and Chair, Computational Imaging, School of Biomedical Engineering and Imaging Sciences, King's College London, UK
When: May 22, 9:45-11 am
Title: Deep learning for smart medical imaging
Abstract: Deep learning approaches in medical imaging have shown great promise in the areas of detection, segmentation and disease classification and are now moving into more complex topics such as motion correction and shape modelling. However, their success is limited by the availability and quality of the images in the dataset used for training these algorithms. A common approach is to train deep learning methods on a well annotated and curated database of high-quality image acquisitions, which then may fail on real patient cases in a hospital setting. In this talk I will show some of our recent deep learning approaches that aim to overcome some of these challenges, by applying novel methods for image augmentation and image compounding. To illustrate some of these approaches, I will draw from examples in cardiac magnetic resonance imaging and fetal ultrasound imaging.
Biography: Julia Schnabel is Professor of Computational Imaging at the School of Biomedical Engineering and Imaging Sciences, King’s College London. She joined King’s in 2015 from the University of Oxford, where she was Professor of Engineering Science. She previously held postdoc positions at University College London, King’s College London and University Medical Center Utrecht. Her research is focusing on machine/deep learning, nonlinear motion modelling, as well as multi-modality, dynamic and quantitative imaging for a range of medical imaging modalities and applications. She is the Director of the Centre for Doctoral Training in Smart Medical Imaging at King’s and Imperial College London, a Director of the Medical Imaging Summer School (MISS), has been Program Chair of MICCAI 2018, General Chair of WBIR 2016, and will be General Co-Chair of IPMI 2021. She is an Associate Editor of IEEE Transactions on Medical Imaging and Transactions on Biomedical Engineering, is on the Editorial Board of Medical Image Analysis, and is an Executive Editor of the new Journal of Machine Learning for Biomedical Imaging (melba-journal.org). She serves on the IEEE EMBS AdCom, the MICCAI Society Board, and has been elected Fellow of the MICCAI Society and Fellow of the European Laboratory for Learning and Intelligent Systems (ELLIS).
Souptik Barua, PhD- Postdoctoral Research Fellow, Scalable Health Lab, Rice University, Houtson, TX
When: May 8th, 9:45 - 11:00 am
NYC location: Virtual only
Ithaca Location:Virtual only
Title: Leveraging structure in cancer imaging to predict clinical outcomes
Abstract: In this talk, I present data-driven frameworks that leverage different types of structure in cancer imaging data to predict clinical outcomes of interest. I demonstrate my findings using two kinds of cancer image data: multiplexed Immuno-Fluorescent (mIF) images from the field of pathology, and Computed Tomography (CT) from radiology. In mIF images, I show that spatial structure based on cell proximities can be used as a visual signature of immune infiltration. Further, the spatial proximity of certain cell types is independently associated with clinical outcomes such as overall survival and risk of progression in pancreatic and lung cancer. In CT images acquired at multiple time points, I demonstrate that the temporal evolution of image features can be used to predict clinical outcomes such as the likelihood of complete response to radiation therapy and the risk of developing long-term radiation injuries such as osteoradionecrosis. Towards the end of my talk, I will present some new research directions in leveraging structure from sensor data in diabetes and pediatric arrhythmias.
Biography: Souptik Barua is a postdoctoral research associate in the Electrical and Computer Engineering department at Rice University. As part of the Scalable Health labs at Rice, Souptik’s research draws on ideas from machine learning, computer vision, and statistics, to discover clinically meaningful information from sensor data. His current focus is on discovering computational biomarkers in diabetes, cancer, and cardiac arrhythmias.
Souptik obtained his Bachelors in Electrical Engineering (B.Tech) from the Indian Institute of Technology, Kharagpur, India in 2012. He received his M.S and Ph.D. in Electrical Engineering from Rice University in 2015 and 2019 respectively. Souptik was one of 13 final-year Ph.D. students invited to the inaugural EPFL Ph.D. summit at Lausanne, Switzerland. He is also a current recipient of a $25k Innovation Seed grant from the NSF as part of the PATHS-UP program.
Ulas Bagci, PhD - Principal Investigator and Assistant Professor, Center for Research in Computer Vision, University of Central Florida, Orlando, FL
When: March 5th, 3:15 - 4:30 pm
NYC location: Belfer (413 E69 St), BB 204-C
Ithaca Location: Weill Hall 226
Title: A Collaborative Computer Aided Diagnosis (C-CAD) System with Eye-Tracking, Sparse Attentional Model, and Deep Learning
Abstract: Vision researchers have been analyzing behaviors of radiologists during screening to understand how and why they miss tumors or misdiagnose. In this regard, eye-trackers have been instrumental in understanding visual search processes of radiologists. However, most relevant studies in this aspect are not compatible with realistic radiology reading rooms. In this talk, I will share our unique experience for developing a paradigm shifting computer aided diagnosis (CAD) system, called collaborative CAD (C-CAD), that unifies CAD and eye-tracking systems in realistic radiology room settings. In other words, we are creating artificial intelligence (AI) tools that get benefits from human cognition and improve over complementary powers of AI and human intelligence. We first developed an eye-tracking interface providing radiologists with a real radiology reading room experience. Second, we proposed a novel computer algorithm that unifies eye-tracking data and a CAD system. The proposed C-CAD collaborates with radiologists via eye-tracking technology and helps them to improve their diagnostic decisions. The proposed C-CAD system has been tested in a lung and prostate cancer screening experiment with multiple radiologists. More recently, we also experimented brain tumor segmentation with the proposed technology leading to promising results. In the last part of my talk, I will describe how to develop AI algorithms which are trusted by clinicians, namely “explainable AI algorithms". By embedding explainability into black box nature of deep learning algorithms, it will be possible to deploy AI tools into clinical workflow, and leading into more intelligent and less artificial algorithms available in radiology rooms.
Biography: Dr. Bagci is a faculty member at the Center for Research in Computer Vision (CRCV), His research interests are artificial intelligence, machine learning and their applications in biomedical and clinical imaging. Previously, he was a staff scientist and the lab co-manager at the NIH's Center for Infectious Disease Imaging (CIDI) Lab, department of Radiology and Imaging Sciences (RAD&IS). Dr. Bagci had also been the leading scientist (image analyst) in biosafety/bioterrorism project initiated jointly by NIAID and IRF. Dr. Bagci obtained his PhD degree from Computer Science, University of Nottingham (UK) in collaboration with University of Pennsylvania. Dr. Bagci is senior member of IEEE and RSNA, and member of scientific organizations such as SNMMI, ASA, RSS, AAAS, and MICCAI. Dr. Bagci is the recipient of many awards including NIH's FARE award (twice), RSNA Merit Awards (5+ times), best paper awards, poster prizes, and several highlights in journal covers, media, and news. Dr. Bagci was co-chair of Image Processing Track of SPIE Medical Imaging Conference, 2017, and technical committee member of MICCAI for several years.
Ben Glocker, PhD - Reader in Machine Learning for Imaging, Faculty of Engineering, Department of Computing, Imperial College London, London UK
When: February 14th, 9:45 - 11:00 am
NYC location: Belfer Building (413 E69), BB 204-C
Ithaca Location: Phillips Hall, 233
Title: Causality Matters in Medical Imaging
Abstract: We use causal reasoning to shed new light on key challenges in medical imaging: 1) data scarcity, which is the limited availability of high-quality annotations, and 2) data mismatch, whereby a trained algorithm may fail to generalize in clinical practice. We argue that causal relationships between images, annotations, and data-collection processes can not only have profound effects on the performance of predictive models, but may even dictate which learning strategies should be considered in the first place. Semi-supervision, for example, may be unsuitable for image segmentation - one of the possibly surprising insights from our causal considerations in medical image analysis. We also discuss two approaches for tackling the problem of domain (or acquisition) shift. We conclude that it is important for the success of machine-learning-based image analysis that researchers are aware of and account for the causal relationships underlying their data.
Biography: Dr. Ben Glocker is Reader (eq. Associate Professor) in Machine Learning for Imaging at Imperial College London. He holds a PhD from TU Munich and was a post-doc at Microsoft and a Research Fellow at the University of Cambridge. His research is at the intersection of medical image analysis and artificial intelligence aiming to build computational tools for improving image-based detection and diagnosis of disease.
Konrad Kording, PhD - Professor, Department of Bioengineering and Neuroscience, University of Pennsylvania, Philadelphia, PA
When: January 23rd, 3:15 - 4:30 pm
NYC location: ST8A-05 (Starr building, floor 8A)
Ithaca Location: Weill Hall 224
Title: Is most of medical machine learning wrong or misleading?
Abstract: The promise to convert large datasets into medical insights is driving the transition of medicine towards a data rich discipline. Consequently, many scientists focus on machine learning from such datasets. Countless papers are exciting, but very little has clinical impact. Here I argue that this is due to the way we do machine learning, and how common practices lead to non-replication or misleading interpretations of machine learning results. I will discuss ways of minimizing such problems.
Biography: Dr. Kording's (He/Him) is trying to understand how the world and in particular the brain works using data. Early research in the Kording lab focused on computational neuroscience and in particular movement. But as the approaches matured, the focus has more been on discovering ways in which new data sources as well as emerging data analysis can enable awesome possibilities. The current focus is on Causality in Data science applications - how do we know how things work if we can not randomize? But we are also very much excited about understanding how the brain does credit assignment. The kording lab style of working is transdisciplinary, we collaborate on virtually every project.
Joaquin Goni, PhD - Assistant Professor, School of Industrial Engineering & Weldon School of Biomedical Engineering, Purdue University, West Lafayette, IN
When: December 6th, 9:45 - 11:00 am
NYC location: Belfer Building, 204-A
Ithaca Location: Phillips Hall 233
Title: Brain connectomics: from maximizing subjects identifiability to disentangling heritable and environment traits
Abstract: In the 17th century, physician Marcello Malpighi observed the existence of patterns of ridges and sweat glands on fingertips. This was a major breakthrough and originated a long and continuing quest for ways to uniquely identify individuals based on fingerprints. In the modern era, the concept of fingerprinting has expanded to other sources of data, such as voice recognition and retinal scans. It is only in the last few years that technologies and methodologies have achieved high-quality data for individual human brain imaging, and the subsequent estimation of structural and functional connectivity. In this context, the next challenge for human identifiability is posed on brain data, particularly on brain networks, both structural and functional.
Here I present how the individual fingerprint of a human structural or functional connectome (as represented by a network) can be maximized from a reconstruction procedure based on group-wise decomposition in a finite number of orthogonal brain connectivity modes. By using data from the Human Connectome Project and from a local cohort, I also introduce different extensions of this work, including an extended version of the framework for inter-scanner identifiability, evaluating identifiability on graph theoretical measurements and an ongoing extended version of the framework towards disentangling heritability and environmental brain network traits.
Biography: I am a Computational Neuroscientist who works in the emergent research area of Brain Connectomics. I am the head of the CONNplexity Lab, which focuses on the application of Complex Systems approaches in Neuroscience and Cognitive Science, including frameworks such as graph theory, information theory or fractal theory. Projects include relating structural and functional connectivity within the human brain. My interest includes healthy and disease conditions, including neurodegenerative diseases. I also make contributions to theoretical foundations of Complex Systems.
I earned my degree in Computer Engineering in 2003 (University of the Basque Country) and my Ph.D in 2008 from the Department of Physics and Applied Mathematics (University of Navarra). After a first postdoc at a Functional Neuroimaging Lab at University of Navarra, from 2011 to 2014 I was a postdoctoral researcher at the group of Dr. Olaf Sporns at Indiana University. In 2015, I joined Purdue University as an Assistant Professor.
Bratislav Misic, PhD - Assistant Professor, Neurology and Neurosurgery, McGill University, Montreal, CA
When: November 15th, 9:45 - 11:00 am
NYC location: Belfer Building (413 E69 St), 302-D
Ithaca Location: Weill Hall, 224
Title: Signaling and transport in brain networks
Abstract: The complex network spanned by millions of axons and synaptic contacts acts as a conduit for both healthy brain function and for dysfunction. Collective signaling and communication among populations of neurons supports flexible behaviour and cognitive operations. Perturbations, such as stimulation-induced dynamic activity or the accumulation of pathogenic proteins, often spread from their source location via axonal projections. Here I will focus on how two fundamental types of dynamics - electrical signaling and molecular transport - can be modeled in brain networks.
Biography: Dr. Bratislav Misic leads the Network Neuroscience Lab. We investigate how cognitive operations and complex behaviour emerge from the connections and interactions among brain areas. The goal of this research is to quantify the effects of disease on brain structure and function. Our research program emphasizes representations and models that not only embody the topological organization of the brain, but also capture the complex multi-scale relationships that link brain network topology to dynamic biological processes, such as neural signalling and disease spread. Our research lies at the intersection of network science, dynamical systems and multivariate statistics, with a focus on complex data sets involving multiple neuroimaging modalities, including fMRI, DWI, MEG/EEG and PET.