Upcoming Seminars

Speaker

Kushagra Chandak, PhD student at the University of Alberta, supervised by Dr. Xiaoqi Tan

Title

Offline-to-Online Learning in Linear Bandits

Abstract
We study online learning with an additional offline dataset in the stochastic linear bandit setting. Although this problem arises frequently in practice, the offline-to-online tradeoff remains poorly understood in structured environments. We propose a linear bandit algorithm that balances this tradeoff: it relies on offline data during early rounds, and increasingly favors exploration as the horizon grows. We establish regret bounds showing that our method is simultaneously competitive with both purely online and purely offline solutions. In particular, it achieves sublinear regret relative to the optimal action in the number of online interactions, while its regret relative to an offline reference decreases as the number of offline samples grows. Empirical results further demonstrate its effectiveness across various problem parameters.

Presenter Bio
Kushagra Chandak is a PhD student at the University of Alberta supervised by Dr. Xiaoqi Tan, where he also received his MSc with Dr. Csaba Szepesvari and Dr. Nidhi Hegde. His research interests are in bandits and reinforcement learning with a focus on how to improve online learning using offline data in terms of efficiency and safety.

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

Add event

Speaker

Laura Petrich, PhD student at the University of Alberta, supervised by Dr. Matthew Taylor and Dr. Patrick Pilarski

Title

Dendritic Gated Networks for Robust Prosthetic Control under Data Distribution Shifts

Abstract
High abandonment rates for upper-limb prostheses are often driven by a loss of user trust when control systems fail to handle the variations of daily life. While current clinical standards like linear discriminant analysis (LDA) provide steady-state reliability, they struggle to adapt to real-world signal shifts caused by muscle fatigue or electrode movement. This seminar presents the use of dendritic gated networks (DGNs), a neural network architecture designed for robust, context-aware intent prediction directly on wearable edge devices.

To address clinical data scarcity, we developed an anatomically-informed protocol to simulate muscle activation patterns, allowing for faster testing and validation of new control models. This talk details offline benchmarking showing that DGNs outperform LDA in dynamic conditions, particularly during electrode shifts and different limb positions, while improving safety by correctly identifying wrong predictions. Furthermore, we show results from real-time, human-in-the-loop virtual reality experiments where DGN-based control achieved higher task success rates and greater subjective reliability than the clinical standard. Finally, we outline future work exploring DGNs for continual learning on hardware with strict resource constraints and their extension to other robotic environments.

Presenter Bio
Laura is currently pursuing a Ph.D. in Computing Science under the supervision of Dr. Patrick Pilarski (BLINC lab) and Dr. Matthew E. Taylor (IRL lab). She received a B.Sc. with Honors in Computing Science from the University of Alberta in 2019 and an M.Sc. in Computing Science from the University of Alberta in 2022. Her research interests include neuroprosthetics, machine learning, and human-robot interaction. Drawing inspiration from her anatomical studies, Laura’s research aims to develop robot control methods with the goal of increased functionality, reliability, and safety in the real-world.

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

Add event

Speaker

Dr. Russ Greiner, Amii Fellow and Professor in the Department of Computing Science

Title

TBA

Abstract
TBA

Presenter Bio
TBA

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

TBA

Speaker

Montaser Mohammedalamen, Applied Research Scientist in Trust and Safety at Amii

Title

TBA

Abstract
TBA

Presenter Bio
TBA

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

Add event

Speaker

Parham Mohammad Panahi, PhD student at the University of Alberta, supervised by Dr. Adam White and Dr. Michael Bowling

Title

TBA

Abstract
TBA

Presenter Bio
TBA

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

TBA

Speaker

Dr. Alasdair Newson, Associate Professor at Sorbonne Université, hosted in Edmonton by Dr. Martha White

Title

Generative Models for Image and Video Inpainting and Editing, and Steering of Multimodal Large Language Models

Abstract
In this talk, I will discuss three topics: video inpainting, image editing and steering of Multimodal Large Language Models (MLLMs). The first two subjects employ the impressive capabilities of deep generative neural networks, in particular Diffusion Models and Flow Matching, which I will also briefly introduce.

The first work focuses on video inpainting, that is, filling in missing or damaged regions in videos. We propose an approach based on diffusion models, a family of particularly powerful generative models that rely on the progressive inversion of a noising process. Unfortunately, the large size of these models makes them difficult to use, especially for high-dimensional data such as videos. Therefore, in this work, we propose a ""frugal"" approach to video inpainting, based on the assumption of self-similarity in videos, i.e., the presence of redundant content. We obtain results that are equivalent or superior to state-of-the-art models, while using much more compact models (by one or more orders of magnitude). The associated source code is available here: https://github.com/ncherel/infusion.

In the second part, I will present a series of works which employ generative models for image editing, with the goal of modifying semantic content of images, such as the expression of a face. These methods structure the latent spaces of generative models to correspond to the desired semantic attributes, enabling simple and controllable editing.

Finally, I will discuss the steering of MLLMs. These are models which can take different modalities such as images, text and audio, and produce some output, often text. Unfortunately, this output may be considered dangerous or undesirable, for example giving unfounded financial advice. The task of guiding the internal representation of a fixed MLLM, such that it produce safe outputs, is referred to as ""steering"". In this work, we train a very small Multi-Layer Perceptron to predict an input-dependent steering shift vector. We show that it indeed increases the safety of the responses in popular LLMs such as LLaVa and Qwen2. 

Presenter Bio
Alasdair Newson has been an Associate Professor at the ISIR (Institut des Systèmes Intelligents et de Robotique) of Sorbonne Université since 2023. He received his PhD from Télécom Paris in 2014, under the supervision of Andrés Almansa, Yann Gousseau, and Patrick Pérez. From 2014 to 2018, he was a postdoctoral researcher at Duke University with Guillermo Sapiro, Université Paris Descartes with Bruno Galerne and Julie Delon, and Télécom Paris with Yann Gousseau and Saïd Ladjal. From 2018 to 2023, he was Assistant Professor at Télécom Paris. His research interests include image and video restoration, editing and analysis, and the study of generative models for these purposes.

Website

Alasdair Newson

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

Add event