You are warmly invited to the next PGR Seminar.

Date & Time: Monday 20/10/2025 14:00-15:00

Location: JC 1.33A

1. Speaker: Erdem Kus

Title: Frugal Algorithm Selection for Combinatorial Search

Abstract: Solvers for combinatorial search and optimisation problems often exhibit highly complementary performance: instances that are hard for one solver may be easy for another. The Algorithm Selection Problem (ASP) addresses this by predicting, for each problem instance, which solver will perform best. Machine learning models trained for this purpose, however, are typically expensive to construct, as they require exhaustive solver runs on all training instances to obtain ground-truth performance data.

In this work, we propose a frugal alternative that formulates algorithm selection as an active learning problem. Instead of uniformly evaluating all solver–instance pairs, our method intelligently selects the most informative ones, thereby drastically reducing the cost of data collection. We show that standard active learning techniques are inadequate for this setting, as they overlook the structure and cost characteristics unique to algorithm selection. To address this, we introduce novel, cost-aware active learning strategies that leverage auxiliary models to balance informativeness and evaluation cost.

Bio: Erdem is a PhD candidate whose research focuses on Artificial Intelligence (AI) and Constraint Programming (CP).

2. Speaker: Junyu Zhang

Title: Remaking Characters in Heritage Contexts to Support Inclusive Learning

Abstract: Characters in immersive environments have the potential to enrich user experience, improving engagement with heritage and in so doing benefiting heritage organisations and their communities. Creating authentic digital scenes based upon survey, archaeological and historical data, co-creative design and community engagement enables communities and their visitors to understand the past better. The understanding of authenticity stimulates the potential of enriching cultural heritage with the details of lives past and also discusses how this research benefits the Sustainable Development Goals.

Bio: Minty is a PhD candidate exploring the authenticity of characters to support inclusive learning in heritage contexts. She is interested in how digital technologies can be used in the intersection of different disciplines to achieve SDGs in the field of cultural heritage, so as to enhance the promotion, representation, and well-being in digital humanities education and also affect resonated dialogue and thinking among diverse people and communities in facing the current challenges.

We hope you can join us!

You are warmly invited to the next PRG Seminar.

Date & Time: Monday 13/10/2025 14:00-14:40

Location: JC 1.33A

Speaker: Qurat ul ain Shaheen

Title: A Framework for Uncertainty Sampling in Active Learning

Abstract: Uncertainty sampling is an active learning paradigm where data instances representing maximum uncertainty for a machine learning model are selected for training. This talk will explore existing uncertainty modelling approaches for binary classification of categorical data.  It will introduce a conceptual framework to improve uncertainty modelling and present some preliminary results.

Bio: Qurat ul ain Shaheen is a final year PhD researcher. Her research focuses on modelling uncertainty in active learning.

We hope you can join us!

You are warmly invited to the next PRG Seminar.

Date & Time: Monday 06/10/2025 14:00-14:40

Location: JC 1.33A

Speaker: David Morrison

Title: Synthetic Whole Slide Image Patch Embeddings for Multiple Instance Learning

Abstract: Obtaining high-quality data is a persistent challenge for the training of computational pathology models. As medical data, Whole-slide images (WSIs) are often held under restrictive terms by medical institutions and, as a result, are hard to access by researchers. Where data is available, the number of whole slide images can be limited and skewed towards common pathology types. In addition, there can be issues with labelling: slide-level labels may lack information about specific pathologies, for example, they may be limited to binary labels of normal or malignant, while annotations at the level of patches are rarely available.

Synthetic data generation is a possible solution to these problems by allowing researchers to produce data on demand that can be used in an unrestricted manner with high-quality labels. I have previously presented on the generation of synthetic patch data. In this talk, I will discuss an extension to this work in which this approach is combined with models trained to characterise the slide as a whole in order to provide a synthesis process for data for use with multiple instance learning techniques, commonly used in whole slide image classification.

We hope you can join us!

 

The next PGR seminar is taking place this Friday 30th May at 2PM in JC 1.33a

Below are the Titles and Abstracts for Lina and Victor’s talks – Please do come along if you are able.

Lina Hadji-Kyriacou

Title: Context-PEFT: Efficient Cross-Domain Transfer Learning

Abstract: Parameter-Efficient Fine-Tuning (PEFT) techniques such as LoRA, BitFit and IA3 have demonstrated comparable performance to full fine-tuning of pre-trained models for specific downstream tasks, all while demanding significantly fewer trainable parameters and reduced GPU memory consumption. However, in the context of cross-domain transfer learning, the need for architectural modifications or full fine-tuning often becomes apparent. To address this we propose Context-PEFT, which learns different groups of adaptor parameters based on the current input domain. This approach enables LoRA-like weight injection without requiring additional architectural changes. Our method is evaluated on the COCO captioning task, where it outperforms full fine-tuning under similar data constraints while simultaneously offering a substantially more parameter-efficient and computationally economical solution.

Victor Yuan

Title: Methodologies for Creating Interactive and Lifelike Historical Characters Based on MetaHuman

Abstract: Virtual characters have long held promise as pedagogical tools in heritage education, particularly for creating immersive interactions with historical figures. Researchers have envisioned systems capable of emulating these figures, enabling users to engage in life-like, face-to-face dialogues over time. While technological constraints historically limited such applications, recent advancements in computational graphics and language models have now made them viable. This paper presents a framework for constructing interactive virtual character systems, outlining their core components through two critical dimensions: photorealism and interaction. The photorealism dimension leverages modern graphics tools to achieve high-fidelity visual representation, while the interaction dimension utilizes language models to enable socially believable and contextually responsive dialogue. We examine the necessity of each component and analyze available technological solutions with their respective trade-offs. Beyond the technical framework, we discuss potential future improvements and address ethical and practical concerns inherent to such systems. By synthesizing current technologies and their applicability, this work provides institutions with practical guidance for developing customized interactive systems that balance functionality with cost-efficiency.

The next PGR seminar is taking place this Friday 23rd May at 2PM in JC 1.33a

Below are the Titles and Abstracts for Thomas and Charis’ talks – Please do come along if you are able.

Thomas Martin

Title: From Isolated to Continuous Automated Sign Language Recognition

Abstract: Sign languages are full-fledged visual natural languages combining manual and non-manual features. As the name suggests, Automated Sign Language Understanding (ASLU) aims to automate tasks involving sign language. A primary obstacle to ASLU is the creation of appropriate datasets. Indeed, most datasets focus on materials gathered from TV broadcasts covering limited topics, which fail to accurately reflect sign language in the wild. Moreover, annotating such datasets is a prohibitively costly process. With the end goal of Sign Language Translation (written/spoken language to sign language) in mind, ASLU research has transitioned from Isolated to Continuous Sign Language Recognition. However, sign language intricacies have made this transition non-trivial.

Charis Hanna

Title: Enhancing Deep Learning Approaches for the Automated Monitoring of Dense Seabird Colonies

Abstract: Cliff-nesting birds serve as valuable indicators of marine ecosystem health, yet dense populations and remote habitats present significant challenges for automated monitoring. With current state-of-the-art object detectors often failing under the conditions of extreme crowding and occlusion, this project aims to develop and refine deep learning techniques that enable the fine-grained, automated analysis of seabird colonies. Current work explores semi-supervised learning strategies that leverage domain-shifted knowledge to reduce the need for exhaustive annotation across complex image datasets. These methods not only reduce the laborious process of manual annotation but also demonstrate promising improvements in performance across the long-tailed species distribution. While ongoing efforts are directed at further optimising these models, future work will leverage additional spatial information with the aim of supporting richer insights into behavioural dynamics within these populations.