The next PGR seminar is taking place this Friday 21st March at 2PM in JC 1.33a

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

Constantine Theocharis

Title: Efficient Programs with Dependent Types

Abstract:

Dependent types allow us to program using the full power of set theory at our disposal. We can encode conditions of arbitrary complexity, and then show that these conditions are met by our programs, statically. While this paradigm is very effective for verifying systems, often their real-world implementations are done in languages without these verification capabilities, because they produce more efficient programs. In this talk, I will explore some of the main sources of inefficiency in (functional) languages with dependent types, and some work that aims to mitigate these, so that verification and implementation can happen in the same language. A common pattern in these languages is to have ‘refinements’ of data which carry along with them proofs of the properties we care about. The first piece of work is about how to make these refinements true zero-cost abstractions. Another source of inefficiency is that these languages must heap allocate almost everything since the sizes of types cannot always be known at compile time. The second piece of work is about how to keep track of type sizes as part of the type system, so that all heap allocations are explicit and unnecessary for the most part.

Yigit Yazicilar

Title: Automated Nogood-Filtered Fine-Grained Streamlining

Abstract:
We present an automated method to enhance constraint models through fine-grained streamlining, leveraging nogood information from learning solvers. This approach reformulates the streamlining process by filtering streamliners based on nogood data from the SAT solver CaDiCaL. Our method generates candidate streamliners from high-level Essence specifications, constructs a streamliner portfolio using Monte Carlo Tree Search, and applies these to unseen problem instances. The key innovation lies in utilising learnt clauses to guide streamliner filtering, effectively reformulating the original model to focus on areas of high search activity. We demonstrate our approach on the Covering Array Problem, achieving significant speedup compared to the state-of-the-art coarse-grained method. This work not only enhances solver efficiency but also provides new insights into automated model reformulation, with potential applications across a wide range of constraint satisfaction problems.

The next PGR seminar is taking place this Friday 14th March at 2PM in JC 1.33a

Below are the Title and Abstract for Mirza’s talk – Please do come along if you are able.

Title: BioFuse: Optimizing Biomedical Embeddings with Foundation Models

Abstract: Pre-trained foundation models have revolutionized biomedical AI, excelling in specialized domains like radiology and histopathology. However, integrating multiple models remains a challenge due to compatibility and feature fusion issues. BioFuse is an open-source framework designed to optimize biomedical embeddings by automatically selecting and fusing the best model combinations. Leveraging 9 state-of-the-art foundation models and a grid search strategy, BioFuse generates task-specific embeddings that improve downstream classification. On the MedMNIST+ benchmark, it achieves SOTA AUC in 5/12 datasets while maintaining near-SOTA performance in others. Surprisingly, our experiments reveal strong cross-modal capabilities, where models trained on one modality perform well on others. With a high-level API and an extensible architecture, BioFuse streamlines model integration and paves the way for new insights in biomedical data fusion.

The next PGR seminar is taking place this Friday 28th February at 2PM in JC 1.33a

Below are Titles and Abstracts for Ben and Erdem’s talks – Please do come along if you are able.

Ben Claydon

Title: Mechanisms for Similarity Search

Abstract:

Similarity search encompasses the task of finding those objects in a large collection which are most alike to, in some way, an object presented by the user as a query. The domain of these objects is wide, from images to text to chemical structures. This task becomes yet harder when the database becomes extremely large, and a sublinear query time with respect to the database size becomes a requirement. This talk discusses why the problem becomes so hard when presented with complex data, and how algorithms and data structures can be engineered to serve these queries.

Erdem Kus

Title: Frugal Algorithm Selection

Abstract: When solving decision and optimisation problems, many competing algorithms (model and solver choices) have complementary strengths. Typically, there is no single algorithm that works well for all instances of a problem. Automated algorithm selection has been shown to work very well for choosing a suitable algorithm for a given instance. However, the cost of training can be prohibitively large due to running candidate algorithms on a representative set of training instances. In this work, we explore reducing this cost by choosing a subset of the training instances on which to train. We approach this problem in three ways: using active learning to decide based on prediction uncertainty, augmenting the algorithm predictors with a timeout predictor, and collecting training data using a progressively increasing timeout. We evaluate combinations of these approaches on six datasets from ASLib and present the reduction in labelling cost achieved by each option.