Date/Time : 24 Nov 2025 14:00
Venue : S17 #04-05 (Seminar Room 2)
Speaker : Wang, Chao, Southern University of Science and Technology

Variational Bayesian inference offers a powerful framework for tackling key challenges in tensor analysis and image restoration, which are critical problems in machine learning and computer vision. This talk presents innovative approaches to address these challenges by integrating Bayesian principles, variational inference, and physical modeling. First, we introduce a Bayesian framework for Tensor Robust Principal Component Analysis (TRPCA) to recover low-rank structures and characterize sparse noise in mixed-noise scenarios. By embedding a low-rank tensor nuclear norm prior and a generalized sparsity-inducing prior within a Bayesian framework, our method learns the optimal tensor nuclear norm and balances low-rank and sparse components automatically. Second, we propose a novel physics-informed optimization framework for jointly solving depth estimation and image restoration from a single defocused image. By modeling the defocused image as a function of a depth map and an all-in-focus (AiF) image based on optical physics, the framework leverages their intrinsic connections. The depth map guides AiF image recovery, while the AiF image regularizes the depth map reconstruction via reconstruction error. A variational inference algorithm, parameterized with deep neural networks, ensures flexibility and high performance.

Add to calendar :

Date/Time : 21 Nov 2025 16:00
Venue : S17 #04-06 (Seminar Room 1)
Speaker : Matthew Shunshi Zhang, University of Toronto

We introduce a new method for analyzing midpoint discretizations of stochastic differential equations (SDEs), which are frequently used in Markov chain Monte Carlo (MCMC) methods for sampling from a target measure π \propto exp(−V). Borrowing techniques from Malliavin calculus, we compute estimates for the Radon-Nikodym derivative for processes which may anticipate the Brownian motion, in the sense that they may not be adapted to the filtration at the same time. Applying these to various popular midpoint discretizations, we are able to improve the regularity and cross-regularity results in the literature on sampling methods. We also obtain a query complexity bound of O(d^{1/4}) for obtaining a ε^2-accurate sample in 𝖪𝖫 divergence, under log-concavity and strong smoothness assumptions for ∇^2 V.

Add to calendar :

Date/Time : 19 Nov 2025 16:00
Venue : S17 #05-11 (Seminar Room 5)
Speaker : Srecko Durasinovic, Nanyang Technological University

Mixture models, such as Gaussian mixture models (GMMs), are widely used in machine learning to represent complex data distributions. A key challenge, especially in high-dimensional settings, is to determine the mixture order and estimate the mixture parameters. We study the problem of approximating a target measure, available only through finitely many of its moments, by a mixture of distributions from a parametric family (e.g., Gaussian, exponential, Poisson), with approximation quality measured by the 2-Wasserstein ($\was$) or the total variation ($\tv$) distance. Unlike many existing approaches, the parameter set is not assumed to be finite; it is modeled as a compact basic semi-algebraic set. We introduce a hierarchy of semidefinite relaxations with asymptotic convergence to the desired optimal value. In addition, when a certain rank condition is satisfied, the convergence is even finite and recovery of an optimal mixing measure is obtained. We also present an application to clustering, where our framework serves either as a stand-alone method or as a preprocessing step that yields both the number of clusters and strong initial parameter estimates, thereby accelerating convergence of standard (local) clustering algorithms.

Add to calendar :

Date/Time : 15 Oct 2025 14:00
Venue : S17 #04-05 (Seminar Room 2)
Speaker : Vladimir Spokoiny, Weierstrass Institute for Applied Analysis and Stochastics (WIAS)

The recent paper Gao et al. (2023) on estimation and inference for top-ranking problem in Bradley-Terry-Luce (BTL) model presented a surprising result: component-wise estimation and inference can be done under much weaker conditions on the number of comparison then it is required for the full dimensional estimation. The present paper revisits this finding from completely different viewpoint. Namely, we show how a theoretical study of estimation in sup-norm can be reduced to the analysis of plug-in semiparametric estimation. For the latter, we adopt and extend the general approach from Spokoiny (2025b) to high-dimensional estimation and inference. The main tool of the analysis is a theory of perturbed marginal optimization when an objective function depends on a low-dimensional target parameter along with a high-dimensional nuisance parameter. A particular focus of the study is the critical dimension condition. Full-dimensional estimation requires in general the condition n ≫ p between the effective parameter dimension p and the effective sample size corresponding to the smallest eigenvalue of the Fisher information matrix . Inference on the estimated parameter is even more demanding: the condition n ≫ p^2 cannot be generally avoided; see Spokoiny (2025b). However, for the sup-norm estimation, the critical dimension condition can be reduced to n ≥ C log p.

Add to calendar :

Date/Time : 30 Sep 2025 10:00
Venue : S17 #06-11 (Seminar Room 6)
Speaker : Haiyun He, Hong Kong University of Science and Technology, Guangzhou

Watermarking has become a critical technique for distinguishing AI-generated text from human-authored content. In this talk, we introduce a novel theoretical framework for both zero-bit and multi-bit watermarking for Large Language Models (LLMs), aiming to jointly optimize both the watermarking scheme and the detection process. Our approach rigorously characterizes the fundamental trade-offs between watermark detectability, text quality, and information embedding rate. We establish the concept of distributional information embedding, where watermarking actively modifies the token generation process rather than embedding signals into a pre-existing text. Our theoretical analysis reveals that the maximum achievable watermarking rate is dictated by the entropy of the LLM’s output distribution and increases with higher allowable distortion.

Leveraging these insights, we propose an efficient, model-agnostic, and distribution-adaptive watermarking algorithm that optimally balances detectability and text quality while maintaining strict false alarm control. Our method employs a surrogate model along with the Gumbel-max trick to achieve superior detection performance. Empirical evaluations on LLaMA2-13B and Mistral-8×7B demonstrate the effectiveness of our approach. Additionally, we explore robustness considerations, paving the way for watermarking systems that can withstand adversarial attacks more effectively.

Add to calendar :

Date/Time : 25 Sep 2025 15:00
Venue : S17 #05-12 (Seminar Room 4)
Speaker : Lars Nerger, Alfred Wegener Institute

Data assimilation is used to estimate the state in components of Earth system models, and to estimate parameters that control processes in these models. The estimate is computed by combining the model state with real observations, like the temperature of the ocean surface or winds in the atmosphere, accounting for the uncertainties in both the model and the observations. The observations are characterized by significant uncertainties and by extreme sparseness. For example, for the ocean satellites can only observe the surface while measurements inside the ocean can only be done by ships or autonomous instruments, of which only a few thousand are distributed over the world ocean. Skillful data assimilation methods have been developed to utilize the sparse observational information with complex models used on supercomputers. Here, ensemble simulations provide uncertainty estimates, but multiply the modeling cost. In addition, the model nonlinearity induces challenges because common methods like ensemble Kalman filters assume Gaussian error distributions. This presentation will discuss the computational, algorithmic, and practical challenges and recent developments in these complex applications of data assimilation.

Add to calendar :

Date/Time : 21 Aug 2025 16:00
Venue : S17 #04-06 (Seminar Room 1)
Speaker : Liu Sheng, Stanford University

While modern machine learning models and systems, including most advanced large language models, have achieved impressive results, their success is still predicated on massive collections of human-curated, carefully filtered data. This reliance exposes a fundamental vulnerability: when deployed in the real world, these models encounter data that is noisy, incomplete, or ambiguous, leading to unpredictable and unreliable behavior. This challenge is more urgent in high-stakes domains such as medicine, where robustness and trustworthiness are paramount.
In this talk, I will present how we can move beyond the limitations of curated data by developing machine learning algorithms and theory that enable models to learn and adapt directly from the complexity of real-world environments. I will share new insights into the underlying mechanisms of failure for current models under imperfect data, and introduce principled methods that address these vulnerabilities. Finally, I will illustrate how these advances empower reliable AI systems for critical medical applications, including early Alzheimer’s disease detection and automated radiotherapy treatment planning.

Add to calendar :

Date/Time : 04 Jul 2025 16:00
Venue : S17 #04-04 (Seminar Room 3)
Speaker : Ye Jieping, Alibaba

The rapid advancement of large language models (LLMs) showcases unprecedented capabilities, yet their opaque internal decision-making remains a critical challenge. Mechanistic interpretability has emerged as a critical frontier for enhancing the safety, reliability, and controllability of LLMs. In this talk, we synthesize cutting-edge research that reverse-engineers LLM internals to establish foundational principles and deployable applications. Through mechanistic analysis of diverse capabilities—including arithmetic computation, multilingual processing, and factual recall—we reveal how LLMs represent and process information as discrete circuits. These insights enable novel applications: a) Precise Control: Token-, activation-, neuron- and module-level intervention to adjust LLM’s behaviour; b) Precise Tuning: Head- and neuron-level tuning to enhance LLM’s performance c) Precise Quantization: Interpretability-guided quantization strategies achieving 3.4-bit parity with 4-bit baselines.

Add to calendar :

Date/Time : 06 Jun 2025 14:00
Venue : S17 #04-05 (SR2)
Speaker : Xie Xingyu, National University of Singapore

Large language models excel at pattern completion yet struggle with multi-step mathematics, logical deduction, and tool-assisted problem solving. Recent work shows that reinforcement learning can provide the “slow-thinking” mechanisms needed for such tasks. This talk formulates reasoning as a sequential decision process and explains how policy-gradient objectives are adapted to text generation. The main case study, DeepSeek-R1, is examined in detail, covering its Group Relative Policy Optimisation objective, reward design, and the resulting gains on contest-level mathematics. The talk also surveys subsequent refinements that build on R1’s principles, illustrating how practical adjustments to reward formulation and output control further enhance reasoning performance. Together these methods illustrate how RL is currently advancing the reasoning abilities of state-of-the-art language models.

Add to calendar :