http://talks.cam.ac.uk/show/rss/5408

Understanding human intent is fundamental to robots that can collaborate naturally and effectively. Intent prediction involves forecasting time-series data – such as human motion trajectories, gaze patterns, and interaction data – to enable machines to anticipate human actions, respond appropriately, and learn from interaction. This capability paves the way for safer, faster, and intuitive human-robot collaboration. This work presents a framework that combines Imitation Learning techniques with Foundation Models to advance intent-aware robot learning. The approach is demonstrated across diverse tasks, including target prediction in extended reality, human-robot handovers, and multi-robot coordination. By leveraging multimodal cues—such as hand motion, gaze, and interaction history – the system enhances prediction accuracy. Additionally, large language and vision models enable the interpretation of high-level human instructions for task planning and robot navigation. Together, these contributions move toward the goal of Physical AI, where robots can learn from humans and understand and act on their intent in real-world environments.

• Speaker: Speaker to be confirmed
• Tuesday 05 August 2025, 12:00-13:00
• Venue: Cambridge University Engineering Department, JDB Seminar Room.
• Series: Probabilistic Systems, Information, and Inference Group Seminars; organiser: Melanie Ellwood.

Add to your calendar or Include in your list

Abstract not available

• Speaker: Christopher Clarke, University of Bath
• Tuesday 21 October 2025, 16:00-17:00
• Venue: Computer Lab, FW26 and Online.
• Series: Mobile and Wearable Health Seminar Series; organiser: Cecilia Mascolo.

Add to your calendar or Include in your list

Modern XR devices can achieve high frame rate, high resolution, wide color gamut, and brighter displays in order to achieve perceptual realism. However, these requirements lead to significant challenges in terms of battery life, with current commercial devices only lasting up to 3 hours on a single charge. Many techniques have been proposed to modulate rendered pixel intensities in order to save display power, but it is unclear how these techniques impact image quality. In this talk, I will describe several techniques for display power reduction, including traditional techniques as well as methods which target wide field of view, stereoscopic displays, and present results from psychophysical studies which can motivate the use of certain algorithms for specific display types.

Join Zoom Meeting ID: 87269487419 Passcode: 774254

• Speaker: Kenny Chen, New York University
• Thursday 31 July 2025, 14:00-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

High Dynamic Range (HDR) imaging captures a wider luminance range, enhancing visual realism beyond standard imaging constraints. We present deep learning-based approaches for HDR reconstruction, generation, and display. For reconstruction, we leverage implicit neural fields with a physics-driven camera model to recover HDR content from multi-focus, multi-exposure image stacks, effectively handling misalignment and depth variations. Additionally, we utilize 3D Gaussian Splatting to reconstruct HDR radiance fields, enabling real-time, depth-of-field cinematic rendering. In HDR generation, we introduce an unsupervised generative model (GlowGAN) that learns HDR distributions from LDR images via exposure-consistent projections, alongside LEDiff, a diffusion-based framework that fine-tunes Stable Diffusion on a limited HDR dataset for HDR synthesis and LDR -to-HDR restoration through latent space fusion. For display, we propose a self-supervised tone mapping approach that optimizes contrast perception at test time, surpassing traditional methods in perceptual fidelity.

Bio: Chao Wang received his Ph.D. from the Max Planck Institute for Informatics, where he was advised by Prof. Karol Myszkowski and Prof. Hans-Peter Seidel. He also worked closely with Prof. Ana Serrano and Dr. Thomas Leimkühler. Before joining MPI , he earned an M.S. degree from Peking University and a B.S. degree from the University of Electronic Science and Technology of China. Chao has gained research experience across academia and industry. He served as a Research Scientist Intern at Adobe’s Next Cam Lab, collaborating with Marc Levoy and Cecilia Zhang, and previously worked as an Algorithm Research Scientist Intern at Tencent’s Robotics X Lab. He was also a visiting researcher at the University of Calgary.

Join Zoom Meeting ID: 81513876666 Passcode: 991761

• Speaker: Dr. Chao Wang, Max Planck Institute for Informatics
• Tuesday 29 July 2025, 14:00-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

Abstract not available

• Speaker: Eyal Kolman (Microsoft)
• Friday 24 October 2025, 12:00-13:00
• Venue: Hybrid (In-Person + Online). Here is the Zoom link: https://cam-ac-uk.zoom.us/j/4751389294?pwd=Z2ZOSDk0eG1wZldVWG1GVVhrTzFIZz09.
• Series: NLIP Seminar Series; organiser: Suchir Salhan.

Add to your calendar or Include in your list

Protein representation learning (PRL) is crucial for understanding structure-function relationships, yet current sequence- and graph-based methods fail to capture the hierarchical organization inherent in protein structures. We introduce Topotein, a comprehensive framework that applies topological deep learning to PRL through the novel Protein Combinatorial Complex (PCC) and Topology-Complete Perceptron Network (TCPNet). Our PCC represents proteins at multiple hierarchical levels—-from residues to secondary structures to complete proteins—-while preserving geometric information at each level. TCP Net employs SE(3)-equivariant message passing across these hierarchical structures, enabling more effective capture of multi-scale structural patterns. Through extensive experiments on four PRL tasks, TCP Net consistently outperforms state-of-the-art geometric graph neural networks. Our approach demonstrates particular strength in tasks such as fold classification which require understanding of secondary structure arrangements, validating the importance of hierarchical topological features for protein analysis.

• Speaker: Zhiyu Wang
• Tuesday 29 July 2025, 17:00-18:00
• Venue: Computer Laboratory, William Gates Building, Lecture Theatre 1.
• Series: Foundation AI; organiser: Pietro Lio.

Add to your calendar or Include in your list

Abstract not available

• Speaker: Gérard Ben Arous (Courant)
• Thursday 12 March 2026, 16:30-17:30
• Venue: Centre for Mathematical Sciences MR2.
• Series: Peter Whittle Lecture; organiser: Richard Samworth.

Add to your calendar or Include in your list

The Weisfeiler-Leman (WL) algorithms form a family of incomplete approaches to the graph isomorphism problem. They recently found various applications in algorithmic group theory and machine learning. In fact, the algorithms form a parameterized family: for each k ∈ ℕ there is a corresponding k-dimensional algorithm WLk. The algorithms become increasingly powerful with increasing dimension, but at the same time the running time increases. The WL-dimension of a graph G is the smallest k ∈ ℕ for which WLk correctly decides isomorphism between G and every other graph. In some sense, the WL-dimension measures how difficult it is to test isomorphism of one graph to others using a fairly general class of combinatorial algorithms. Nowadays, it is a standard measure in descriptive complexity theory for the structural complexity of a graph. We prove that the WL-dimension of a graph on n vertices is at most 3/20 ⋅ n o(n) = 0.15 ⋅ n o(n). Reducing the question to coherent configurations, the proof develops various techniques to analyze their structure. This includes sufficient conditions under which a fiber can be restored uniquely up to isomorphism if it is removed, a recursive proof exploiting a degree reduction and treewidth bounds, as well as an exhaustive analysis of interspaces involving small fibers. As a base case, we also analyze the dimension of coherent configurations with small fiber size and thereby graphs with small color class size.

• Speaker: Pascal Schweitzer (TU Darmstadt)
• Wednesday 23 July 2025, 11:00-12:00
• Venue: Computer Laboratory, William Gates Building, Room GS15.
• Series: Algorithms and Complexity Seminar; organiser: Tom Gur.

Add to your calendar or Include in your list

As a form of computer graphics, vector graphics represents visual content based directly on geometric shapes (via command lines and arguments), and is widely used in scientific and artistic applications, including architecture, surveying, 3D rendering, typography, and graphic design. Compared with raster graphics, vector graphics is preferred when a high degree of geometric precision is required across arbitrary scales, among which Scalable Vector Graphics (SVG) is a popular vector graphics file format extensively employed in creative industries. Generally, creating SVG content is difficult for non-professional users. It is tedious and time-consuming to gain adequate knowledge of SVG grammars and/or master professional editing software such as Adobe Illustrator. This seminar aims to provide a basic introduction to how recent advances in deep learning are used to improve vector‑graphics generation, animation, and vectorization.

Zoom link: https://cam-ac-uk.zoom.us/j/89875598428?pwd=rzzh5WOnK49HSeaO0I2Wsbz7EgDFbB.1

• Speaker: Ronghuan Wu, City University of Hong Kong
• Thursday 24 July 2025, 14:00-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

As a form of computer graphics, vector graphics represents visual content based directly on geometric shapes (via command lines and arguments), and is widely used in scientific and artistic applications, including architecture, surveying, 3D rendering, typography, and graphic design. Compared with raster graphics, vector graphics is preferred when a high degree of geometric precision is required across arbitrary scales, among which Scalable Vector Graphics (SVG) is a popular vector graphics file format extensively employed in creative industries. Generally, creating SVG content is difficult for non-professional users. It is tedious and time-consuming to gain adequate knowledge of SVG grammars and/or master professional editing software such as Adobe Illustrator. This seminar aims to provide a basic introduction to how recent advances in deep learning are used to improve vector‑graphics generation, animation, and vectorization.

• Speaker: Ronghuan Wu, City University of Hong Kong
• Thursday 24 July 2025, 14:00-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

Modern XR devices can achieve high frame rate, high resolution, wide color gamut, and brighter displays in order to achieve perceptual realism. However, these requirements lead to significant challenges in terms of battery life, with current commercial devices only lasting up to 3 hours on a single charge. Many techniques have been proposed to modulate rendered pixel intensities in order to save display power, but it is unclear how these techniques impact image quality. In this talk, I will describe several techniques for display power reduction, including traditional techniques as well as methods which target wide field of view, stereoscopic displays, and present results from psychophysical studies which can motivate the use of certain algorithms for specific display types.

• Speaker: Kenny Chen, New York University
• Thursday 31 July 2025, 14:00-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

Modern XR devices can achieve high frame rate, high resolution, wide color gamut, and brighter displays in order to achieve perceptual realism. However, these requirements lead to significant challenges in terms of battery life, with current commercial devices only lasting up to 3 hours on a single charge. Many techniques have been proposed to modulate rendered pixel intensities in order to save display power, but it is unclear how these techniques impact image quality. In this talk, I will describe several techniques for display power reduction, including traditional techniques as well as methods which target wide field of view, stereoscopic displays, and present results from psychophysical studies which can motivate the use of certain algorithms for specific display types.

• Speaker: Kenny Chen, New York University
• Thursday 31 July 2025, 14:00-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

Abstract not available

• Speaker: Meytar Ronel, Department of Veterinary Medicine
• Friday 10 October 2025, 08:45-10:00
• Venue: LT2.
• Series: Friday Morning Seminars, Dept of Veterinary Medicine; organiser: Fiona Roby.

Add to your calendar or Include in your list

Abstract not available

• Speaker: Tobias Roddiger, Karlsruhe Institute of Technology (KIT).
• Tuesday 18 November 2025, 16:00-17:00
• Venue: Computer Lab, FW26 and Online.
• Series: Mobile and Wearable Health Seminar Series; organiser: Cecilia Mascolo.

Add to your calendar or Include in your list

Abstract not available

• Speaker: VP Nguyen, , UMass Amherst
• Tuesday 04 November 2025, 16:00-17:00
• Venue: Online.
• Series: Mobile and Wearable Health Seminar Series; organiser: Cecilia Mascolo.

Add to your calendar or Include in your list

The increasing availability of human trajectory and social data, fueled by GPS and social networks, presents a unique opportunity for scientific discovery. However, existing data analysis methods struggle to provide robust, efficient, and generalizable graph representations, hindering their applicability in urban and social sciences. This research addresses this challenge by developing novel machine learning algorithms specifically tailored for graph-structured data in these domains. This research tackles three key challenges: (1) Sparse Data and Data Distribution Heterogeneity: Current methods often struggle with sparse data and varying data distributions, limiting their ability to capture diverse patterns and hindering scalability. This research proposes novel approaches for flexible, adaptive, and generalizable representations in urban planning and social sciences. (2) Non-General Representation and Difficulty Adapting to New Data: Existing methods often lack the ability to generalize across different datasets and struggle to adapt to new data, hindering their effectiveness in real-world applications. This research aims to develop methods that can learn robust and efficient representations that generalize across different datasets and adapt to new data. (3) Trade-off Between Efficiency and Effectiveness: Balancing processing speed, accuracy, and reliability is crucial in urban and social science data analysis. This research addresses this challenge by developing innovative algorithms that optimize for both efficiency and effectiveness. This research leverages contrastive learning and information bottleneck techniques to develop robust and efficient graph representation learning methods for spatial-temporal data and recommender systems. The developed methods have demonstrated significant improvements in downstream tasks such as traffic prediction, crime prediction, and anomaly detection. This research lays a strong foundation for future work in graph-structured data analysis across various domains, including urban science, social science, and scientific discovery. Future research will focus on extending these methods to multi-modal datasets, enabling zero-shot learning, and developing novel approaches for understanding complex biological systems. The Zoom link is shown as follows: https://hku.zoom.us/j/92081548742?pwd=mvQFsCafLqNHGPySmT7isKgxp0aEmH.1

• Speaker: qrzhang@hku.hk
• Monday 28 July 2025, 15:50-17:20
• Venue: Computer Laboratory, William Gates Building, Lecture Theatre 1.
• Series: Foundation AI; organiser: Pietro Lio.

Add to your calendar or Include in your list

We have seen remarkable growth in smart devices and artificial intelligence in all aspects of our lives. Despite these successes, there still exists a large gap between realizing robust, practical, and omnipresent AI that coexists and interacts with humans and the physical world. There are many great success stories in bringing the benefits of AI from the productivity tools, digital assistants, and chatbots in the digital world to our physical world, ranging from smart wearables that track health and fitness, to sensors for monitoring the environment and robots that interact with our physical world. Traditionally, these diverse areas are explored by equally diverse communities. In this talk, we argue that to create truly autonomous and intelligent physical spaces, we need to co-design applications and services at all scales and levels with both humans and computers in the loop, ranging from infrastructure found all around our environments (e.g., sensors and smart speakers) down to wearables and personal smartphones enable humans and computers to interface with each other. Towards this vision, we will present several lines of work that 1) demonstrates how co-optimizing our environments with humans-in-the-loop can significantly improve important metrics for both our built environments (e.g., energy consumption) and humans (e.g., health and comfort), 2) enables new modalities for computers and AI to interact with our physical world through drones, robots, and foundation models, and 3) enhances the natural language interface between humans and the digital world through efficient AI architectures for speech enhancement.

Biography: Stephen Xia is an assistant professor in the Department of Electrical and Computer Engineering at Northwestern University. His research lies at the intersection between systems, embedded machine learning, and signal processing, spanning areas in mobile and embedded computing, Internet-of-Things, cyber-physical systems, artificial intelligence, and smart health. His work focuses on realizing truly intelligent and autonomous environments by embedding and utilizing compute, perception, actuation, storage, and networking resources commonly found all around us. Stephen’s research has been highlighted by many popular media outlets, including Mashable, Fast Company, and Gizmodo, and has received various distinctions including multiple Best Demo Awards, Best Presentation Awards, and Best Paper Awards. Prior to Northwestern, Stephen was a postdoctoral scholar at UC Berkeley, received his Ph.D. from Columbia University, and received his B.S. from Rice University, all in Electrical Engineering.

https://cam-ac-uk.zoom.us/j/82323711212?pwd=jrpvWLn5A57FGIVOAVSJ5Gv0lVLERB.1

• Speaker: Stephen Xia, Northwestern University
• Wednesday 30 July 2025, 14:00-15:00
• Venue: Computer Lab, FW26 and Online.
• Series: Centre for Mobile, Wearable Systems and Augmented Intelligence Seminar Series; organiser: Cecilia Mascolo.

Add to your calendar or Include in your list

The increasing availability of human trajectory and social data, fueled by GPS and social networks, presents a unique opportunity for scientific discovery. However, existing data analysis methods struggle to provide robust, efficient, and generalizable graph representations, hindering their applicability in urban and social sciences. This research addresses this challenge by developing novel machine learning algorithms specifically tailored for graph-structured data in these domains. This research tackles three key challenges: (1) Sparse Data and Data Distribution Heterogeneity: Current methods often struggle with sparse data and varying data distributions, limiting their ability to capture diverse patterns and hindering scalability. This research proposes novel approaches for flexible, adaptive, and generalizable representations in urban planning and social sciences. (2) Non-General Representation and Difficulty Adapting to New Data: Existing methods often lack the ability to generalize across different datasets and struggle to adapt to new data, hindering their effectiveness in real-world applications. This research aims to develop methods that can learn robust and efficient representations that generalize across different datasets and adapt to new data. (3) Trade-off Between Efficiency and Effectiveness: Balancing processing speed, accuracy, and reliability is crucial in urban and social science data analysis. This research addresses this challenge by developing innovative algorithms that optimize for both efficiency and effectiveness. This research leverages contrastive learning and information bottleneck techniques to develop robust and efficient graph representation learning methods for spatial-temporal data and recommender systems. The developed methods have demonstrated significant improvements in downstream tasks such as traffic prediction, crime prediction, and anomaly detection. This research lays a strong foundation for future work in graph-structured data analysis across various domains, including urban science, social science, and scientific discovery. Future research will focus on extending these methods to multi-modal datasets, enabling zero-shot learning, and developing novel approaches for understanding complex biological systems. The Zoom link is shown as follows: https://hku.zoom.us/j/92081548742?pwd=mvQFsCafLqNHGPySmT7isKgxp0aEmH.1

• Speaker: qrzhang@hku.hk
• Monday 21 July 2025, 15:50-17:20
• Venue: Computer Laboratory, William Gates Building, Lecture Theatre 1.
• Series: Foundation AI; organiser: Pietro Lio.

Add to your calendar or Include in your list

As a form of computer graphics, vector graphics represents visual content based directly on geometric shapes (via command lines and arguments), and is widely used in scientific and artistic applications, including architecture, surveying, 3D rendering, typography, and graphic design. Compared with raster graphics, vector graphics is preferred when a high degree of geometric precision is required across arbitrary scales, among which Scalable Vector Graphics (SVG) is a popular vector graphics file format extensively employed in creative industries. Generally, creating SVG content is difficult for non-professional users. It is tedious and time-consuming to gain adequate knowledge of SVG grammars and/or master professional editing software such as Adobe Illustrator. This seminar aims to provide a basic introduction to how recent advances in deep learning are used to improve vector‑graphics generation, animation, and vectorization.

• Speaker: Ronghuan Wu, City University of Hong Kong
• Thursday 24 July 2025, 14:30-15:00
• Venue: SS03 - William Gates Building.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

Abstract not available

• Speaker: Kaiwen Zuo
• Tuesday 29 July 2025, 16:00-17:00
• Venue: Computer Laboratory, William Gates Building, Lecture Theatre 1.
• Series: Foundation AI; organiser: Pietro Lio.

Add to your calendar or Include in your list