In recent years, multiparameter persistent homology (MPH) has been developed to address the limitations of traditional single-parameter persistence. While much progress has been made on the theoretical aspects of MPH, practical applications have mostly been limited to the special case of two parameters. Building on recent developments in the computation of MPH in more than two parameters, we propose and implement a new algorithm for the computation of multiparameter persistence landscapes in this setting. We apply this work to compute landscapes of three-parameter persistence modules built on spatiotemporal data arising in biology.
\r\n
The talk is jointly organized by the EPFL AI Center and the EPFL Machine Learning for Biomedicine Lab (MLBIO) as part of the AI Fundamentals seminar series.
\r\n
\r\nTitle
\r\nAgents That Know What They Know: Interpretability and Introspection for Lifelong Learning
\r\n
\r\nAbstract
\r\nOne of the hallmarks of intelligence is the ability to learn from experience. In artificial intelligence, we frame this as lifelong learning: autonomous agents that continue to adapt after deployment as they encounter new tasks and environments. In this talk, I will argue that successful lifelong learning agents require a form of introspection: they must be able to reason about the scope and limits of their own knowledge. This self-knowledge allows an agent to answer three fundamental questions. What do I know? When can I reuse what I know? And what do I still need to learn? I will present recent work from my group showing how ideas from interpretability and explainability can surface the knowledge needed to answer these questions. I will first show how learned representations can be used to identify when inputs are novel, enabling out-of-distribution detection. Then, I will discuss how this can be built upon to give agents a way to distinguish between situations where prior knowledge can be safely reused and situations where new learning is needed. Together, this work shows that interpretability is not only a tool for transparency, but a mechanism for building agents that recognize, reuse, and acquire knowledge over time.
\r\n
\r\nBio
\r\nJoseph Campbell is an Assistant Professor in the Department of Computer Science at Purdue University, where he leads the Collaborative AI for Machines and People Lab. His research bridges machine learning and robotics, with a focus on explainable machine learning and lifelong learning. Before joining Purdue, he was a Postdoctoral Fellow in the Robotics Institute at Carnegie Mellon University, and he received his Ph.D., M.S., and B.S. degrees from Arizona State University. His research is supported by Amazon, the Toyota Research Institute, and the National Science Foundation.
Many species live in groups and coordinate behavior with conspecifics, but the sensory computations that allow individuals to detect, evaluate, and respond to others remain difficult to study. A major challenge is that social interactions are reciprocal: animals continuously change each other’s sensory experience, making it hard to separate stimulus, perception, internal state, and behavioral response.
\r\n
\r\nIn my lab, we use juvenile zebrafish shoaling as an experimentally accessible model for vertebrate social behavior. Using virtual reality, we previously identified self-like biological motion as a visual trigger of shoaling and traced this signal into a tecto-thalamic pathway that contributes to social attraction. We now build on this framework to ask how elementary visual cues are combined into richer social representations.
\r\n
\r\nI will discuss two examples. In one project, behavioral experiments and neural recordings suggest that zebrafish evaluate group size through an integrated visual variable that combines object number and motion, so that faster small groups can become behaviorally equivalent to slower larger groups. In another, we study social buffering: the reduction of fear-related behavior by conspecific cues. Here, social information can either transmit immediate alarm or provide persistent safety, depending on context. Whole-brain activity mapping reveals neural correlates of fear state and its suppression by social cues.
\r\n
\r\nTogether, these projects show how controlled visual stimuli, quantitative behavior, whole-brain activity mapping, and cellular-resolution imaging can reveal how animals transform simple sensory features into social decisions. More broadly, this work aims to understand how nervous systems extract meaning from the motion and presence of others — from detecting social cues to finding safety in numbers.
Thesis Director: Prof. Z. Patakfalvi,
\r\nMathematics doctoral program
\r\nThesis Nr. 12192
\r\n
\r\nTo take part in the public defense, please contact directly the speaker
Thesis Director: Prof. S. C. Olhede,
\r\nMathematics doctoral program
\r\nThesis Nr. 11276
\r\n
\r\nTo take part in the public defense, please contact directly the speaker