Inspiration Exchange is a series of engagement sessions aiming to share ideas and discuss topics that will define the future of machine learning in healthcare. These events will target machine learning researchers interested in healthcare, and will emphasize sharing of new ideas and development of new methods, approaches, and techniques.
As a lab, our purpose is to create new and powerful machine learning techniques and methods that can revolutionize healthcare. This doesn’t happen in a vacuum. At inception, we are inspired by ideas and discussions; in implementation, we need connections, trust, and partnership to make a real difference.
While you can learn about our work at major conferences in machine learning or in our papers, we think it’s a better idea to create a community and keep these conversations going. We’re also aware that many people—both in healthcare and machine learning—have questions about what we do, and how they can contribute.
For more information about Inspiration Exchange—and to sign up to join in—please have a look at the sections below, and keep checking for new updates.
Inspiration Exchange
Themed discussion sessions specifically for machine learning researchers in academia or industry interested in healthcare.
We would like to:
– discuss machine learning models and techniques
– share ideas about how machine learning can revolutionize healthcare
– spark new projects and collaborations
– raise awareness about this unique and exciting area of machine learning.
Standard session format:
– brief tutorial on chosen topic
– presentation of cutting-edge solutions by lab researchers
– roundtable discussions with industry leaders
– Q&A
The next session is on 2 October 2023!
Join us to explore Data-Centric AI, a cornerstone in shaping robust and trustworthy machine learning systems. The quality of data used to train ML models determines the success or failure of AI. Data-centric AI provides a framework that brings data to the fore in designing ML pipelines. This is especially critical given the widespread use of data-hungry algorithms. Despite the importance of “data” nuances and particularities, there is little work that focuses on how to exploit them in our quest for designing robust and reliable ML. Hence, there is an urgent need to shift the focus to the data used in AI/ML and its quality.
We’re eager to share recent advances and discuss new, unexplored directions that could yet provide new breakthroughs. For more info about data-centric AI, visit our website: https://www.vanderschaar-lab.com/data-centric-ai/, and see our previous IE session (with Isabelle Guyon), and, DC-Check, an actionable checklist for practitioners and researchers to elicit data-centric considerations through different stages of the ML pipeline.
Time zones
Cambridge, United Kingdom Mon, June 12, 2023 at 4:00 pm BST
Paris, France Mon, June 12, 2023 at 5:00 pm CEST
New York, USA Mon, June 12, 2023 at 11:00 am EDT
San Francisco, USA Mon, June 12, 2023 at 8:00 am PDT
Beijing, China Mon, June 12, 2023 at 11:00 pm CST
More time zones can be found here
Sign-up form
Session 1: introduction to automated machine learning
Ahead of this session, Mihaela published a piece of content on machine learning and the future of healthcare, entitled AutoML: powering the new human-machine learning ecosystem.
Session 2: automated machine learning pipelines
This session featured presentations from 4 of the van der Schaar Lab’s researchers regarding the Lab’s key AutoML projects. In addition, three session participants presented their own proposals for building on these projects by improving the AutoML methods or applying them in new contexts or applications.
Session 3: software packages for automated machine learning
This session featured presentations from 4 of the van der Schaar Lab’s researchers regarding the lab’s AutoML software packages (AutoPrognosis and Clairvoyance), followed by a Q&A.
Session 4: recent projects in machine learning for healthcare
This session covered a range of the lab’s most recent and exciting research projects in machine learning for healthcare. A total of 9 short presentations were given by the van der Schaar Lab’s researchers, followed by a Q&A session.
Session 5: synthetic data concepts and approaches
This session focused on synthetic data generation as one of the lab’s key research areas. Following an introduction from Mihaela van der Schaar, 3 short presentations were given by the van der Schaar Lab’s researchers, followed by a Q&A session.
Session 6: synthetic data evaluation
This session focused on evaluation of synthetic data. Following an introduction from Mihaela van der Schaar, 2 short presentations were given by the van der Schaar Lab’s researchers, followed by a Q&A session.
Session 7: application-oriented projects in machine learning for healthcare
This session featured 4 presentations (given by current and former lab members) on a range of application-oriented projects in machine learning for healthcare. Topics ranged from treatment effect estimation to multi-omics data integration, organ transplantation, and clinical trials.
A Q&A/open discussion took place in the latter half of the session, with participants asking researchers about their projects, and sharing their thoughts.
Session 8: individualized treatment effect inference (key concepts and challenges)
This was the first of 2 planned sessions focusing on individualized treatment effect inference (ITE inference).
The former half of the session featured 2 presentations introducing individualized treatment effect inference. This was followed by a participative discussion in which members of the lab and session participants discussed several key topics related to ITE inference, ranging from the evaluation of models to working with different types of data, to potentially relaxing some of the typical assumptions related to ITE inference.
Session 9: individualized treatment effect inference (time series)
This was the second in a pair of sessions focusing on individualized treatment effect inference (ITE inference). Building on our discussions the previous month, this session examined a range of ITE inference problems and approaches specifically within the time series setting.
The former half of the session featured 4 short presentations, followed by a participative discussion in which members of the lab and members of the audience discussed several key topics relevant to ITE inference in the time series setting, including counterfactuals, assumptions, missing data, and interpretability.
Session 10: quantitative epistemology
This session offered an introduction to quantitative epistemology—the van der Schaar Lab’s transformational new research area that aims to use AI and machine learning to understand and empower human learning and decision-making.
The former half of the session featured a high-level introduction to quantitative epistemology by Mihaela van der Schaar, followed by more in-depth explanations of specific frameworks and approaches by Ph.D. student Alihan Hüyük. The latter half of the session focused on a participative discussion in which Mihaela and Alihan answered questions from audience members about quantitative epistemology.
Session 11: time series in healthcare
The focus of the session was time series—an absolutely vital topic (and key research pillar for the van der Schaar Lab) that touches all aspects of machine learning for healthcare.
The first half of the session featured short presentations given by Mihaela van der Schaar, former Ph.D. student Changhee Lee, and research engineer Evgeny Saveliev. The latter half of the session featured a Q&A, in which Mihaela and members of the lab answered questions from audience members about time series in healthcare.
Session 12: what’s next for individualized treatment effects?
This session revisited the crucial topic of individualized treatment effect (ITE) inference, with a particular focus on looking to the future and introducing and discussing some brand-new approaches.
The first half of the session featured short presentations given by Ph.D. students Alicia Curth and Zhaozhi Qian. The latter half of the session featured a Q&A, in which Mihaela, Alicia, and Zhaozhi answered questions from audience members about their research and individualized treatment effects in general.
Session 13: how can we make ML models as robust and useful as possible?
This session raised a critically important question: how can we make ML models as robust and useful as possible?
Looking beyond predictions and analytics, there are a range of requirements and attributes that affect the degree of robustness and utility of a machine learning model—both in and beyond healthcare.
This session focused on three of these in particular: synthetic data, uncertainty estimation, and data imputation. Our lab members presented new and compelling work on each of these, while also conducting polls of the audience and asking attendees for their opinions and personal experiences.
Session 14: ML to transform organ transplantation
This was a particularly application-oriented session on ML tools for organ transplantation (one of the lab’s key research impact areas).
To start the session, clinician Alexander Gimson framed the importance and complexity of organ transplantation as a domain, and gave a quick overview of some of his collaborative projects with the lab to date.
Ph.D. student Jeroen Berrevoets then introduced OrganITE and OrganSync, two new ML-powered approaches to organ transplantation problems such as transplant benefit prediction, matchmaking, and queueing. Research engineer Bogdan Cebere presented a demonstrator for OrganBoard, a clinician-oriented survival prediction and matchmaking tool for organ transplantation. Finally, postdoc Fergus Imrie and PhD. student Yuchao Qin presented iTransplant, a machine learning project aiming to understand and improve decision-making around accepting and rejecting organs.
The session wrapped up with a short roundtable with a panel of clinical collaborators (all of whom are transplantation experts) and a brief audience Q&A.
Session 15: next frontiers in interpretability
This was an especially interactive session, featuring an audience poll and a lively Q&A and discussion.
To start the session, Mihaela van der Schaar shared a framework for understanding and categorizing the various different types of machine learning interpretability (in and beyond the healthcare setting).
Ph.D. students Jonathan Crabbé and Zhaozhi Qian then presented two fundamentally different ways of approaching interpretability (and related projects), and discussed their respective strengths. The session was then opened up to the audience for Q&A and general discussion.
Session 16: discovery using machine learning
This was an especially interactive session, featuring an audience poll and a lively Q&A and discussion. The session topic was “Discovery Using Machine Learning.”
To start the session, Mihaela shared her reasons for believing that machine learning can enable new discoveries (both in and beyond the healthcare setting) and why this is a new frontier for machine learning.
Ph.D. students Zhaozhi Qian and Krzysztof Kacprzyk then presented how machine learning can discover governing equations from data, highlighting a recent approach from our lab, and discussed the limitations of previous approaches. The session was then opened up to the audience for Q&A and general discussion.
Session 17: Causal Deep Learning
This was our first Inspiration Exchange session following a newly introduced format. This was an especially interactive session, featuring a roundtable discussion with industry guests, lively Q&A and discussions. The session topic was “Causal deep learning.”
To start the session, Mihaela shared a vision for combining causality and deep learning methods, which was further developed by Ph.D. student Zhaozhi Qian in an extended tutorial, where he presented the “Rung 1.5” framework for causality. Current and former lab members, Ioana Bica and Dr. Alexis Bellot, then presented two ways of adopting causality principles in generalisable imitation learning and causally aware imputation.
We then held our first roundtable discussion, where our industry guests Silvia Chiappa (DeepMind), Kim Branson (GSK), and Lindsay Edwards (Relation Therapeutics) shared their perspectives on practical challenges that can be addressed through causal deep learning. The session was then opened up to the audience for Q&A and general discussion.
Session 18: Causal Deep Learning Part 2
The session was a part 2 discussion on “Causal deep learning.”
To start the session, Ph.D. student Jeroen Berrevoets presented an extended tutorial, showcasing the “Rung 1.5” approach for imputation when estimating treatment effects. Subsequently, Ph.D. student Boris van Breugel introduced a method to adapt causality principles in generating fair synthetic data.
We then held our roundtable discussion, where our industry guests Amit Sharma (Microsoft Research), Cheng Zhang (Microsoft Research), and Jim Weatherall (AstraZeneca) shared their perspectives on practical challenges that can be addressed through causal deep learning.
The session was then opened up to the audience for Q&A and general discussion.
Session 19: ICML 2022 Preview
The session previewed the van der Schaar Lab’s accepted publications at the upcoming ICML 2022 conference and featured a roundtable discussion with industry guests, lively Q&A, and discussions.
To start the session, there was a medley of presentations highlighting our 7 accepted papers, highlighting our lab’s continuing and impactful research. Our research encompasses a diverse array of topics, including discovering diverse classes of differential equations from data, machine learning interpretability, synthetic data, data-centric AI, AutoML, individualised treatment effects, and, last but not least, augmenting human skills using machine learning.
We then held our roundtable discussion, where our industry guests Dr Ari Ercole (University of Cambridge) and Dr Razvan Pascanu (DeepMind) shared their perspectives on ML and the medical impacts of the presented works.
The session was then opened up to the audience for Q&A and general discussion.
Session 20: Neural differential equations
The session covered neural differential equations
To start the session, Ph.D. student Zhaozhi Qian introduced neural differential equations and motivated their use in the context of healthcare and medicine. Subsequently, Ph.D. student Samuel Holt presented Neural Laplace, a unified framework for learning diverse classes of differential equations by modeling dynamics in the Laplace domain. Additionally, Ph.D. student Nabeel Seedat presented Treatment Effect Neural Controlled Differential Equation (TE-CDE), which estimates counterfactual outcomes over time with irregularly sampled data.
We then held our roundtable discussion, where our industry guests Dr Emma Slade (GSK) and Dr Javier Hernández (Microsoft Research) shared their perspectives on ML and the medical impacts of the presented works. The session was then opened up to the audience for Q&A and general discussion.
The session was then opened up to the audience for Q&A and general discussion.
Session 21: Frontiers in ML Interpretability
Recording of the van der Schaar Lab’s 21st Inspiration Exchange session, which took place on October 25, 2022, and was attended by students and professionals from the AI and machine learning community.
To start the session, Prof. Mihaela van der Schaar presented a tutorial and introduction to ML interpretability, motivating their use in the context of healthcare and medicine. Subsequently, Ph.D. student Jonathan Crabbé presented his recent work on Concept Activation Regions, a generalised framework for concept-based explanations. Additionally, Ph.D. student Alicia Curth presented the ITErpretability benchmark, which utilises interpretability as a means to benchmark treatment effect models.
We then held our roundtable discussion, where our industry guests Dr Been Kim (Google Brain) and Dr Jianying Hu (IBM) shared their perspectives on ML and medical impacts of the presented works. The session was then opened up to the audience for Q&A and general discussion.
Sign up for our upcoming sessions: https://www.vanderschaar-lab.com/enga… The lab’s publications are here: https://www.vanderschaar-lab.com/publ…
Session 22: AutoPrognosis 2.0
In this Inspiration Exchange session, we introduced and discussed our lab’s latest tool – AutoPrognosis 2.0!
To start the session, Prof Mihaela van der Schaar gave an introductory presentation about AutoPrognosis 2.0, followed by a demonstration of the tool by Dr Fergus Imrie. Subsequently, Dr Thomas Callender (UCL) presented recent work utilising AutoPrognosis for lung cancer risk prediction. Core of the session was the roundtable discussion, where guests Dr Aditya Nori (Microsoft Health Futures), Prof Eoin McKinney (CCAIM), and Dr Thomas Callender (UCL) shared their perspectives on AutoPrognosis 2.0 and autoML in health. The session was then opened up to the audience for Q&A and general discussion. To learn more about AutoPrognosis 2.0, please see our recorded video, our dedicated website, and our paper.
Sign up for our upcoming sessions: https://www.vanderschaar-lab.com/enga… The lab’s publications are here: https://www.vanderschaar-lab.com/publ…
Session 23: Data-centric AI
This session was particularly engaging and interactive, with more than 120 participants joining the session live!
To start the session, Prof Mihaela van der Schaar gave an introductory presentation about Data-centric AI, followed by a presentation about our group’s work in the field by Dr Fergus Imrie, and a deep dive into Data-IQ by Nabeel Seedat. Core of the session was the roundtable discussion with our guest Prof Isabelle Guyon (Google Brain; INRIA) who shared her perspectives on Data-centric AI, challenges, and opportunities. The session was then opened up to the audience for Q&A and general discussion.
To learn more about Data-Centric AI, please see our recorded video, and see our dedicated website for Data-Check: Data-Centric AI Checklist.
Sign up for our upcoming sessions: https://www.vanderschaar-lab.com/enga… The lab’s publications are here: https://www.vanderschaar-lab.com/publ…
Session 24: Synthetic Data
This session was particularly engaging and interactive, with more than 140 participants joining the session live!
To start the session, Prof Mihaela van der Schaar gave an introductory presentation about synthetic data. This was followed by a presentation by Dr. Zhaozhi Qian about our group’s latest open-source package Synthcity, which features a variety of advanced synthetic data generators.
The core of the session was the roundtable discussion with our panel of experts from healthcare, pharmaceuticals, consulting, and finance industries, including Dimitrios Vlitas (Accenture), Elena Sizikova (FDA), Nick Brown (AstraZeneca), Peter Krusche (Novartis), Robert Tillman (Optum Labs), and Tucker Balch (JP Morgan), who shared their perspectives on Synthcity and synthetic data. The session was then opened up to the audience for Q&A and general discussion.
To learn more about Synthcity, please see our recorded video below, our Github Repo, and our paper. We are also holding a hybrid synthetic data tutorial at AAAI-23 on Feb 8, 2023 at 2-6 PM EST, register here to learn more about Synthcity and synthetic data generation.
Session 25: Causal Deep Learning
In this session, Prof. Mihaela van der Schaar and Ph.D. student Jeroen Berrevoets unveiled our lab’s latest causal deep learning (CDL) manifesto (paper and slides attached). Please find below an abstract of our vision for the future of CDL:
Causality has the potential to truly transform the way we solve a large number of real-world problems. Yet, so far, its potential remains largely unlocked since most work so far requires strict assumptions which do not hold true in practice. To address this challenge and make progress in solving real-world problems, we propose a new way of thinking about causality – we call this causal deep learning. The framework which we propose for causal deep learning spans three dimensions: (1) a structural dimension, which allows incomplete causal knowledge rather than assuming either full or no causal knowledge; (2) a parametric dimension, which encompasses parametric forms which are typically ignored; and finally, (3) a temporal dimension, which explicitly allows for situations which capture exposure times or temporal structure. Together, these dimensions allow us to make progress on a variety of real-world problems by leveraging (sometimes incomplete) causal knowledge and/or combining diverse causal deep learning methods. This new framework also enables researchers to compare systematically across existing works as well as identify promising research areas which can lead to real-world impact.
The session also featured a lively roundtable discussion with our panel of experts Dr. Cheng Zhang (Microsoft), Dr. Guillermo Sapiro (Apple, Duke), and Dr. Simon Woodhead (Eedi), who shared their insights on the applications and developments of causal deep learning.
For more resources on causal deep learning, please see:
- Causal deep learning paper (manifesto paper)
- Jeroen’s presentation (attached)
- Recording of Feb 28 session (Youtube video)
- Research pillar on causal deep learning (research pillar)
- Previous sessions on causal deep learning (part 1, part 2)
We encourage everyone to engage with this agenda, and together build the future of CDL!
Session 26: Revolutionizing Clinical Trials with Machine Learning
To start the session, PhD Student Alihan Hüyük gave an introductory presentation about clinical trials and how machine learning can help to decide when to make and break commitments to a trialing process. The core of the session was the roundtable discussion with our panel of experts, including Dr Carl-Fredrik Burman (AstraZeneca), Dr Miro Dudík (Microsoft Research), Dr Sean Hayes (Merck & Co.), and Prof Suetonia Palmer (University of Otago) sharing their insights on the applications and developments of machine learning for clinical trials. The session was then opened up to the audience for Q&A and general discussion.
For more resources on machine learning and clinical trials, see our research pillar on next-generation clinical trials and our perspective on how machine learning can revolutionize clinical trials.
Session 27: AISTATS & ICLR Preview 2023
As a recap, we had the following presentations:
Nabeel Seedat will present his work on how self-supervised learning can improve the quality of the conformal predictions for uncertainty quantification;
Alihan Hüyük will present Neural Laplace Control – a continuous-time model-based offline reinforcement learning method that combines a Neural Laplace dynamics model with a model predictive control planner;
Tennison Liu will present GOGGLE, a deep generative model for tabular data that exploits relational structure and an end-to-end message passing scheme to generate realistic synthetic tabular data;
Samuel Holt will present deep generative symbolic regression, which discovers closed-form mathematical equations from data by leveraging pre-trained deep generative models.
We encourage you to learn more about the presented work by reading the described papers. The titles, authors, abstract, and perceived impacts of all 11 accepted papers can be found on our lab website.
Session 28: 2nd AISTATS & ICLR Preview 2023
Thanks so much to everyone who joined us at the recent Inspiration Exchange session, presenting our lab’s ICLR/AISTATS publications! You can now find the recording of the session on Youtube.
As a recap, we had the following presentations:
Alicia Curth presented her work on inferring heterogeneous treatment effects from time-to-event data in the presence of competing risks;
Boris van Breugel presented DOMIAS, a density-based membership inference attack model that aims to infer membership from synthetic data by detecting local overfitting of generative models;
Alan Jeffares presented TANGOS, a novel regularization method built on latent neuron attributions that encourages orthogonalization and specialization of neuron attributions to improve generalization performance;
Yuchao Qin presented T-Phenotype, a novel temporal clustering method that discovers phenotypes of predictive temporal patterns from time-series data.
We encourage you to learn more about the presented work by reading the described papers. The titles, authors, abstract, and perceived impacts of all 11 accepted papers can be found on our lab website.
Session 29: Treatment effect estimation and ML for Clinical Trials
As a recap, Alicia Curth presented three papers to be published at ICML 2023. Specifically, she discussed 1) how to conduct model selection in the absence of counterfactual information, 2) accounting for informative sampling when forecasting treatment effects over time, and 3) adaptively identifying populations with treatment benefits in clinical trials.
For more, please see the publications:
On CATE model selection: https://arxiv.org/abs/2302.02923
On informative sampling: https://arxiv.org/abs/2306.04255
On adaptive trials: https://arxiv.org/abs/2208.05844
Additionally, the session was joined by Emre Kiciman (Microsoft Research) and Tom Diethe (AstraZeneca), who shared their insights on the developments in ML for treatment effect estimation and clinical trials.
To learn more about our team of researchers, click here. You can find our publications here.