RHUL CS seminars
Welcome to the RHUL CS seminars webpage. Our seminars run on Wednesdays during term time.
Each seminar is tagged with a general Topic and with Technical if technical content is to be expected or Departmental if it is suitable for a general audience. Joint seminars with Center for AI & Skills is denoted by + Center for AI & Skills. The department also runs a separate 🎂 PhD Cake Talks series for our PGR students.
Click on Show Details to see the venue, link to live-streaming, the abstract, and a link to the recording after the seminar (if available).
| Time | Speaker | Title | |
|---|---|---|---|
| Autumn 2025 | |||
| 05 Nov 2025, 16:00 | Shalini Maiti (Meta AI & UCL) | Deep learning methods for 3D reconstruction and evaluation Machine Learning Technical + Centre for AI & Skills | Show Details |
VenueBedford 0-07 [MS Teams Link] Short BioShalini Maiti studied Information and Communication Technology at DA-IICT in India, did her masters at TU Graz, specializing in Computer Vision and Machine Learning. She is currently in the final year of her PhD at University College London and Meta AI. Shalini works with 3D vision, particularly with reconstruction, evaluation and generation of 3D objects. When she's is not huddled over academic deadlines, she spends time with good stories, shared experiences, pub quizzes and travel. AbstractReconstructing and evaluating 3D content poses challenges at both the modeling and assessment stages. For non-rigid objects, 3D recovery from 2D keypoints is ill-posed due to occlusions and entanglement between viewpoint and shape variation. Classical low-rank models impose global constraints but suffer from alignment difficulties and limited expressivity. By instead constraining localized subsets of shape within high-capacity unsupervised models, it is possible to preserve flexibility while ensuring geometric consistency, yielding over 70% error reduction on S-Up3D. In parallel, the rapid growth of text-to-3D generation has exposed limitations of current evaluation metrics, which either require ground-truth supervision (e.g., PSNR) or only measure prompt fidelity (e.g., CLIP). Gen3DEval addresses this by leveraging vision-language models fine-tuned for 3D object quality assessment, enabling reference-free evaluation across text fidelity, appearance, and surface geometry. Together, these approaches advance both the generative and evaluative foundations of 3D vision, highlighting pathways toward more accurate, scalable, and human-aligned 3D understanding. Seminar Recording |
|||
| 29 Oct 2025, 12:30 | Alessandro Pierro (Intel & LMU Munich) | Hardware-Algorithm Co-Design for ML Inference Machine Learning Technical + Centre for AI & Skills | Show Details |
VenueMoore Annex 034B [MS Teams Link] Short BioAlessandro Pierro is a researcher at Intel Labs and a doctoral candidate in computer science at the Ludwig-Maximilians-Universität München. His research focuses on accelerating machine learning and mathematical optimization workloads through hardware-algorithm co-design, using advances in parallel computing architectures. He is currently working on energy-efficient inference for linear recurrent networks to enable sequence modeling at the edge. AbstractThe growing demand for AI accelerators presents an opportunity to validate the technological readiness of emerging hardware architectures. This requires understanding how alternative computational paradigms perform on real applications and identifying which algorithmic trends align with their inherent strengths. This seminar will provide empirical results on current ML workloads running on the Intel Loihi 2 accelerator, a sparse, event-driven, spatially-mapped system. Results on State Space Models and recurrent LLMs demonstrate where Loihi 2 can excel, as well as which architecture-level enhancements could improve its performance on these workloads. We will also cover the implications for hardware architectures arising from recent trends in large-scale ML workloads, including sparse mixtures-of-experts, recurrent reasoning, speculative decoding, and hierarchical networks. Seminar Recording |
|||