Upcoming Seminars

NO SEMINAR - see you next week!

Speaker

Dr. Jonathan Schaeffer

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. Xi Wang, Assistant Professor, University of Sheffield, UK, hosted by Dr. Eleni Stroulia

Title

Beyond Passive Processing: Advancing Proactiveness and Interpretability in Conversational AI

Abstract
As Large Language Models (LLMs) transition into autonomous agents, conversational systems must proactively guide human interactions while remaining internally transparent and verifiable. This talk bridges these twin dimensions by exploring how user-facing proactive strategies intersect with internal model mechanics. We first address strategic interaction through COPE, a Mixture-of-Experts architecture evaluated on our fine-grained InPE dataset. By separating decision pathways via a hierarchical hard-routing scheme, COPE demonstrates that optimal preference elicitation is stage-dependent, dynamically transitioning from abstract, attribute-based inquiries to concrete, item-based evaluations as user intent matures.  However, as proactive systems rely more on external knowledge, ensuring response reliability is paramount. Shifting focus to model internals, we introduce ARC-JSD (Attributing Response to Context via Jensen-Shannon Divergence), an inference-time context attribution mechanism that localises supporting information without heavy fine-tuning or gradient calculations. By pairing this bounded metric with Logit Lens probing, we map the specific attention heads and multilayer perceptron (MLP) layers responsible for context tracking, leveraging them as confidence gates to actively mitigate hallucinations. Together, these frameworks underscore a shift toward conversational agents that are strategically proactive in user interactions, yet rigorously controllable in their internal reasoning.

Presenter Bio
Xi Wang is a Lecturer (~Assistant Professor) in the School of Computer Science at the University of Sheffield, where he leads the Interactive NLP Research Lab (iNLP). His research sits at the cutting edge of Conversational AI and Information Retrieval, with a particular focus on proactive large language models (LLMs), retrieval-augmented generation (RAG), and AI frameworks for healthcare.

He completed his PhD in Computing Science at the University of Glasgow as a member of the Terrier Team under the supervision of Prof. Iadh Ounis and Prof. Craig Macdonald. His doctoral work specialised in advanced personalisation and recommendation frameworks leveraging textual user reviews. Following his time at Glasgow, he joined University College London (UCL) as a Postdoctoral Research Fellow in the Web Intelligence Group, where he collaborated on task-oriented dialogue systems. His research addresses both user-facing strategic interaction and the internal mechanics of language models, with his work regularly published in top-tier venues including ICLR, SIGIR, ACL, EMNLP, and CIKM.

Website

Xi Wang

Timing & Location

UComm Seminar Room 2-108

pizza from 11:30, seminar from noon to 1

Add event to calendar

Add event

Speaker

TBA

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

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

Title

Prediction and Control Learning in Dynamic, Sensory Rich Environments

Abstract
Upper-limb prosthesis abandonment remains high, often driven by a loss of user trust when control systems fail to accommodate the variability of daily life.  While clinical standards like linear discriminant analysis (LDA) offer steady-state reliability, they lack the representational flexibility to manage real-world signal shifts. This thesis proposes the use of dendritic gated networks (DGNs), a biologically-plausible neural network architecture that leverages local learning rules and structural gating to enable robust, context-aware intent prediction.

We address the barrier of clinical data scarcity by developing an anatomically-informed protocol that simulates post-surgical activation patterns in participants without amputation. Comprehensive offline benchmarking across twelve datasets, which encompass confounding factors known to degrade controller performance, demonstrates that DGNs outperform LDA in dynamic conditions. Specifically, results indicate an accuracy improvement of nearly 18% during electrode shifts. Furthermore, DGNs demonstrate superior uncertainty calibration, identifying and rejecting ambiguous predictions that cause overconfident errors in commercially-available prosthetic control models. This can help create prosthetic systems that are more reliable and safe to use.

Building on these results, my proposed work includes a systematic investigation of DGNs for continual learning on prosthetic systems to evaluate their capacity for lifelong intent prediction in dynamic environments. We will benchmark DGNs against established algorithms to evaluate their performance-to-compute ratio in human-in-the-loop contexts. These findings are supported by ongoing virtual reality crossover studies, where preliminary results indicate that DGN-based control facilitates higher functional success rates and reliability of control during complex manipulation tasks. This research further evaluates the scalability of DGNs within the strict power, memory, and latency constraints of wearable edge devices. Beyond the prosthetic domain, we will explore the generalization of dendritic gating to autonomous agents to determine if structural weight partitioning provides a universal advantage for navigating real-time adaptation in complex, sensory-rich robotic environments. Ultimately, this work provides a technical framework for assistive technologies that adapt alongside the individuals they support.

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, usability, 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

TBA

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

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

TBA

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