Human-Centered AI Conference Program

Jonathan Kao

AI for brain-computer interfaces and neuroscience

Abstract: AI is transforming many areas of science and medicine. In this talk, I will give an overview of how AI is advancing brain-computer interfaces, which aim to restore communication and movement to people with paralysis. I will also discuss how AI is giving us new tools to study the inner workings of the brain.

Katie Bouman

Images of the Hidden Universe

Some of the most iconic images in modern science were never captured by a camera in the traditional sense. Instead, they were inferred from indirect and incomplete measurements, using a combination of physics, prior knowledge, and computation. In this talk, I will explore how physics and machine learning are working together to illuminate parts of the universe that are difficult -- or even fundamentally impossible -- to observe directly. I’ll begin with the story of how our Event Horizon Telescope team created the first direct image of a black hole. Theory had long predicted what we should see, and confidence came not from a single image, but from the consistency of features across many reconstructions of the same data. I will discuss how these ideas extend beyond black holes to other scientific imaging problems, including mapping the distribution of dark matter from subtle distortions in the shapes of galaxies due to gravitational lensing. Together, these examples illustrate how modern imaging increasingly relies on integrating physics and machine learning to extract meaningful information from fundamentally limited data to uncover our hidden universe.

Paul Thompson

Can AI Learn to Diagnose Every Brain Disease? Vision-Language Models, Scaling Laws, & Worldwide Neuroimaging

Each year, 30 million MRI scans are acquired in the U.S. alone. Could AI learn from this wealth of data to diagnose every disease of the brain? For the last 15 years, our ENIGMA Consortium (http://enigma.ini.usc.edu) has published the largest brain imaging studies of over 30 brain disorders, in a massive worldwide collaboration where over 2,000 scientists work together to analyze over 300,000 brain scans and genomic data from 45 countries. We study Alzheimer's and Parkinson's disease, epilepsy, stroke, schizophrenia, bipolar disorder, depression, Tourette syndrome, eating disorders, autism and PTSD, as well as other rarer conditions. This makes ENIGMA a proving ground for medical AI on a worldwide scale. This talk tours the models being benchmarked in ENIGMA and beyond: 3D convolutional neural nets and Swin Transformers that detect Alzheimer's disease with ~92% accuracy, explainable-AI maps that reveal what these networks can see, and vision-language models that answer clinical questions about a 3D brain scan in plain language. These models are built by aligning image and text encoders using contrastive pretraining. We then turn to a question facing anyone who trains AI models: how much data is enough? A new "Zeta Law of Discoverability" links diagnostic accuracy to partial sums of the enigmatic Riemann zeta function, a mysterious function that arises in perhaps the greatest unsolved problem of mathematics, the Riemann hypothesis. This scaling law predicts when high-capacity models overtake simpler ones and how we can tell if we just need more data, new kinds of data, better models, or a mixture of these. Finally, we look ahead to decoding mental states from fMRI, generative models of the brain's wiring, and image-driven gene discovery. We also cover the pitfalls of shortcut learning, domain shift and fairness.

David S. Liebeskind, MD, MBA USC Neurovascular Center

Neurovascular AI at the Bedside: Seeing the Mechanism, Choosing the Moment, Preserving the Person

Artificial intelligence offers opportunities to improve neurovascular care by connecting imaging findings with disease mechanisms, treatment decisions, and the patient’s clinical course. This presentation examines established and emerging applications across ischemic and hemorrhagic stroke, intracranial atherosclerosis, aneurysms, vascular malformations, dissection, moyamoya, cerebral venous thrombosis, and small-vessel disease. Clinical CT, MRI, and angiographic examples illustrate how AI may assist detection, quantify tissue injury and perfusion, identify interval change, and support secondary prevention and recovery. Particular attention is given to distinguishing diagnostic performance from evidence that an AI-supported care pathway improves patient outcomes.

Drawing on experience organizing 25 years of UCLA imaging through a stroke census, the presentation outlines a proposed USC program linking neurovascular imaging with longitudinal clinical information across three hospitals using disparate imaging and electronic health record systems. A federated approach would enable collaborative analysis while maintaining local stewardship of source records. Proposed initial studies address recurrent ischemia, injury after acute treatment, and functional recovery, with prospective evaluation at each participating hospital.

The presentation also considers how OpenAI-enabled clinical assistance could evolve toward source-linked record synthesis, imaging comparison, and clinician-reviewed workflows. Throughout, the emphasis remains on clinical accountability, explicit uncertainty, and the distinction between predicting an outcome and identifying an intervention that improves it. Human-centered neurovascular AI should ultimately be evaluated by its contribution to timely decisions, preserved function, and outcomes that matter to patients.

Josh Bloom

AI-accelerated Science will Always be a Human Endeavor

As a data deluge swamped traditional approaches, astronomy (by necessity) turned to AI/ML to facilitate discovery and inference at scale. General-purpose reasoning systems now extend that acceleration beyond the data to the conduct of science itself: synthesizing the literature, writing code, generating hypotheses, and increasingly writing and reviewing scientific publications. I argue that this last step is where the deepest change lies, that the institutions of publishing and peer review were not built for a world in which agents read and write alongside people. However much of the record machines produce, the questions, the trust, and the credit remain fundamentally a human endeavor. I will ground this talk with examples from my own work in real-time transient discovery and physics-informed inference, and from ongoing efforts at Valency where we are creating infrastructure for a scholarly record that people and agents build together.

Leo Petrossian

Restoration, Not Augmentation: What the First FDA-Cleared Brain-Computer Interface Therapy Teaches Us About AI and the Brain

Most conversations about AI and the brain imagine augmentation: a brain learning to control a machine. This talk argues for the opposite future. Kandu markets the first FDA-cleared brain-computer interface therapy, IpsiHand, a non-invasive system in which a machine learns to read a stroke survivor's intent from the uninjured hemisphere and uses it to retrain the brain to move an arm that was written off as permanently lost. The FDA-validated result is restored function across the whole upper extremity: hand, wrist, elbow, and shoulder. Drawing on clinical trials and real-world data from more than a thousand patients treated at home, Leo Petrossian, a serial entrepreneur and CEO in neuroscience, will show what restorative AI looks like in practice today, where it goes next, and why the largest near-term gains for the brain may come from AI applied to the delivery of care rather than to the science of it.

Jie Li

"The Art of Being Human in Age of AI"

AI is getting very good at modeling the world it can see written down: its geometry, physics, and languages. However, it is not getting good at modeling the world that never makes it into text: the flinch a participant doesn’t name in an interview, the exact resistance of buttercream against a spatula that tells your hand it’s ready, the private thought you have and never type. As a researcher trained to sit with people empathetically , and as a self-taught cake artist whose hands have learned things through thousands of repetitions that my words still can’t fully carry, I work in this gap. 

Drawing on a decade of research into how designers, users, and computational systems co-evolve as long-term collaborators, this talk explores the current debate over synthetic user research versus empathetic user research, cake-making as a way to explore embodied knowledge that is learned through touch and practice but difficult to put into words, and our current TeleAbsence project at MIT Media Lab, which insists on recalling a moment rather than recreating it. My argument: synthetic users, world models, and AI companions are all racing to reconstruct human experience from what gets recorded. Most of what makes us human is never recorded in the moment it happens. As a researcher, I don’t see this as a gap in the AI development roadmap, but as the job of a researcher to preserve what remains uniquely human, and of a maker whose hands still know things her words don’t.

Leonardo Christov-Moore

Toward a Robot Boddhisaatva: From Embodiment to Generalization, Care, and Self-Transcendence

Here, we explore work from a neuroscience of empathy perspective, incorporating elements of Heideggerian and Buddhist philosophy, to argue that the inescapable conditions of embodiment, namely:

• vulnerability, the fact of being in the world that one is observing and acting upon
• impermanence/mortality, the fact that there are absorbing states in the world that we are inexorably drawn toward can prove a foundation for generalization in the face of the open-endedness of the physical world, as well as a capacity for care.

Furthermore, via the variable boundary of self posed by embodiment and via intrinsic drives to maintain oneself, the self can be extended to include others. As we move from the era of scaling back to an era of research, such considerations may allow for more efficient and trustworthy AI.

Nick Vincent

Towards a 'Fully Attested' AI supply chain: Developments in AI evaluation and implications for a more human-centered data ecosystem

AI capabilities have continued to improve, leading to increased attention within both research communities and popular discourse on AI evaluation. There is renewed interest in developing better evaluation techniques and developing new kinds of evaluation-focused institutions and coalitions; national news is starting to cover AI safety topics that were previously niche and policy discussions are becoming more intense. In this talk, I'll connect this surge in interest in AI safety to long-running discussions about AI's challenges with data ecosystems. I'll argue that this interest in evaluation represents an opportunity to establish trustworthy AI supply chains, with "attested" evaluation and training data linked to model outputs.