Machine Learning seminars

Robot missions require both collision-free trajectories and logical task ordering, such as collecting a resource before entering a restricted region. Augmented graphs of convex sets combine continuous trajectory optimization and discrete task sequencing in one problem. A layered graph constructed from an exact convex partition of free space selects an optimal task order and trajectory through a shortest-path calculation, exact up to finite Bézier parameterization. Its structure corresponds to Bellman–Held–Karp dynamic programming for the travelling-salesperson problem, retaining singly exponential worst-case complexity and improving exponentially on general temporal-logic tools. A library of specifications covers ordered collection, alternative keys, combined prerequisites, timing and conditional logic with correctness guarantees. Numerical benchmarks demonstrate exponential speedups and near-global optimality. Possible extensions include moving environments, safety-aware planning with conformal prediction sets and a large galactic-survey benchmark.

Jan Kosinski explains how AlphaFold-style methods predict protein complexes and why their reliance on evolutionary information makes interactions between host and pathogen proteins especially difficult. The talk reviews large-scale studies, interpreting both successful predictions and failures rather than treating reported success rates as universal performance. It then examines ways to improve predictions through broader sampling, altered sequence alignments and experimental restraints, including crosslinking mass spectrometry, with influenza A virus as a worked research example. The session is suitable for students and researchers interested in host-pathogen interactions and structural bioinformatics; basic knowledge of protein structures and sequence alignments is useful, but prior AlphaFold experience is unnecessary.

Matthias Vigl examines whether the computational scaling strategies behind modern machine learning can benefit particle physics. Empirical scaling laws relate model performance to computing resources and help balance model size against training-data volume. For a fixed dataset, increasing model capacity beyond the interpolation threshold can improve generalization through double descent, but gains eventually encounter limits imposed by the data. The talk then considers scaling data and model size together, deriving compute-optimal relations for transformer-based jet taggers trained on as many as billions of simulated jets. It studies how training hyperparameters and input representations change these relations. The observed behaviour suggests that larger computing budgets could yield useful gains in particle physics, especially with a move toward more general-purpose foundation models.

Deep networks are often thought of as black boxes. Their ability to encompass vast swathes of knowledge indeed makes them hard to explain. Yet many of their behaviours — generalization under overparametrization, grokking, OOD failures, neural scaling laws — recur across architectures and scales, and each, however surprising, can be reproduced and explained in controlled settings. I will review these efforts to identify and explain the universal phenomena of deep learning, and suggest that an LLM may amount to a sum of such tractable sub-phenomena, interpolative in nature. Finally, leaving scientific rigor aside, I'll argue that what separates this prosaic picture from the apparent magic of LLMs may well be the industrial scale of compute and human labour behind it, and that AGI in its deeper extrapolative sense may be much further away than claimed.

Jie Yang presents BRIDGE, a multilingual evaluation benchmark built from real clinical tasks and more than one million samples derived from electronic health records. The benchmark addresses the gap between simplified examination-style tests and the complexity of clinical data, while enabling comparison of rapidly changing language models. Evaluation of 95 models involved over 24,000 experiments and 39 million predictions. Results vary substantially by model family, task and language, with some open-source systems matching proprietary models. The study also finds that chain-of-thought prompting frequently reduces accuracy on these tasks and examines stigmatizing language produced during model reasoning at large scale.

How can learning systems recommend actions that prevent an undesirable predicted outcome? Rehearsal learning approaches this problem through influence relations, but existing parametric assumptions can limit its use in complex systems. This talk presents a non-parametric method based on conditional mean embeddings. It expresses the objective in a reproducing kernel Hilbert space, replaces a discontinuous desirability indicator with a smooth Probit surrogate, and uses nested kernel ridge regression to estimate outcome distributions conditional on actions. The resulting estimator can be identified from observational data and has finite-sample error bounds and consistency guarantees. Synthetic and semi-synthetic experiments assess its effectiveness and flexibility.

Real-time measurement of local biomolecule concentrations in living tissue requires sensors that are stable, selective and minimally invasive. This seminar examines single-walled carbon nanotubes, whose durable near-infrared fluorescence and sensitivity to their surroundings support optical biosensing. Biopolymer functionalization gives the nanotubes biological compatibility and selectivity for particular molecular targets. Spectroscopy, microscopy and machine-learning-assisted analysis are used to characterize sensor interactions and extract biological information, with applications spanning live-cell imaging, wearable sensing and continuous monitoring.

Tim Ebbels presents interpretable machine-learning models for metabolomics and multi-omics using pathway and chemical-class scores. Sparse coverage of the metabolome and uncertain annotations make conventional results difficult to interpret; pathway-based features aim to connect model output directly to biological function. The webinar examines applications from multi-assay mass spectrometry and multi-omics integration to single-cell and mass-spectrometry imaging, discusses methodological challenges and attempts to resolve them, and demonstrates how these approaches can clarify complex biological signals. Online via Zoom. Free for interested life scientists and bioinformaticians, with advance registration required through the Australian BioCommons event page. Tuesday 13 October, 16:00–17:00 AEDT (Australia/Melbourne), equivalent to 05:00–06:00 UTC. The event is subject to the BioCommons Code of Conduct and may be recorded.

Liver toxicity can emerge late in drug development when early assays fail to capture human biology, exposure or injury mechanisms. This webinar introduces Axiom’s hepatic profiling approach, combining multicellular liver systems, high-content imaging, transcriptomics, proteomics and ADME measurements. Its reasoning agent integrates those results with partner assays, chemical-series context, exposure estimates and clinical evidence to assess mechanistic liver-safety risks. Rather than using a single cell-toxicity score, the approach develops interpretable, series-specific structure–toxicity hypotheses, compares analogues, ranks compounds and suggests follow-up experiments or structural changes. Performance on established liver-injury reference compounds and newer molecules is examined through case studies, showing how multimodal evidence can inform medicinal-chemistry decisions and predicted therapeutic index.

Jin Zhu examines statistical approaches to identifying text produced by large language models. The increasing availability of fluent generated writing raises questions about misinformation, academic integrity and the authenticity of digital content. Although detection has attracted substantial machine-learning research, the talk argues that its statistical foundations remain underdeveloped. Zhu presents recent work on principled, computationally efficient detection methods, including approaches reported at NeurIPS and ICLR. An interactive demonstration accompanies the research, illustrating one of the proposed methods.

Raymundo Arróyave discusses alloy discovery guided by both metallurgical knowledge and Bayesian optimization. The approach combines physical models, high-throughput CALPHAD calculations, microstructural information and experimental feedback to search chemical and processing spaces that are difficult to explore through data alone. Examples involving high-entropy and refractory alloys address multiple objectives and constraints, including feasibility, correlated properties, uncertainty and microstructural sensitivity. The seminar then connects these decision methods to the Autonomous Robotic Metallurgist under development at his university. This platform links modular synthesis, characterization and testing with digital twins, human–robot collaboration and AI agents. The aim is a closed experimental loop in which scientific judgment and physical understanding guide automated execution.

Representation learning underlies visual search, retrieval and recognition by encoding gallery images and matching them with queries. Most early work assumes a fixed model, but new training data or more expressive architectures can change the feature representation. Recomputing every gallery vector after an update is expensive for collections containing billions of images and may be impossible when the original images are unavailable because of privacy or storage restrictions. Compatible representation learning aims to update the model while retaining usable existing gallery features. The talk examines the foundational compatibility problem, the role of representation stationarity, and new training techniques that use stationarity to learn compatible feature representations.

Dan Meyer considers the questions generative AI raises for mathematics learning after four years of rapid technological development. He argues that its novelty has been accompanied by educational adoption and institutional change falling short of early expectations, including lower-than-anticipated teacher use. The session explores what the technology can do, what students need and what teaching entails, using these questions to examine AI’s contribution to education.

Oden Institute Seminar by Rex Ying (Yale University) on how non-Euclidean geometries, particularly hyperbolic geometry, can enhance foundation models by better capturing hierarchies and symmetries in real-world data, with applications across Transformers, language model training, multimodal systems, and recommender systems.

David Persson presents Polar Express, a method for polar decomposition and the matrix sign function motivated by the Muon neural-network optimizer. It uses matrix multiplications suited to GPUs and adapts each polynomial update through a minimax problem to reduce worst-case error. The talk covers convergence, finite-precision implementation in bfloat16 and validation-loss improvements when training GPT-2 on FineWeb data. This is an in-person PACM IDeAS seminar at Princeton.

José Molina Mora presents an integrative study of the core stress-response mechanisms, or perturbome, of Escherichia coli, Pseudomonas aeruginosa and Staphylococcus aureus. Transcriptomic data are analysed using machine learning, systems biology, functional enrichment and ortholog comparisons to prioritise stress-associated genes. Experimentally determined protein structures are combined with AlphaFold predictions where structural data are missing. Molecular docking and further computational analyses then identify preliminary compounds and conditions that might inhibit selected targets. These results are a proof of concept awaiting experimental validation. The talk connects gene-expression analysis with structural modelling for antimicrobial research and is aimed at researchers, graduate students and professionals in microbiology, bioinformatics and related biomedical fields.

Olivia Di Matteo introduces quantum computing and compilation, examining how the translation from high-level quantum programs to efficient hardware instructions becomes a bottleneck as applications scale. The seminar identifies parts of the quantum software stack where machine learning and artificial intelligence can improve this process. It presents the group’s work using graph-based reinforcement learning to optimize quantum circuits, then considers outstanding challenges and research questions for scalable quantum software.

Stephanie Gil studies reliable coordination in robot networks facing malicious information and ordinary environmental uncertainty. Communication signals provide physical evidence of trustworthiness that is difficult to forge; the cy-trust framework turns that evidence into probabilistic trust estimates and weights neighboring agents accordingly. For consensus and distributed optimization, the analysis establishes almost-sure convergence with bounded departures from nominal performance even when malicious agents form a majority of a node’s neighbors, exceeding the classical Byzantine threshold. Theory and hardware experiments support these guarantees. For natural uncertainty, real-time sensing is incorporated into rollout-based reinforcement learning, reweighting possible futures. Applications include fleet routing under random demand and Project CETI’s autonomous robotic rendezvous with sperm whales at sea. The talk closes with directions for combining trust and long-horizon sequential decisions to retain resilience when planning data may be corrupted.