Research Continuity and Project Lineage
Several projects continue completed work, explore a new branch of the same research question, or prepare a shared foundation for a later project. The maps below show those explicit relationships and projects connected by a shared research topic.
How to read the map: Arrows indicate research continuity. Branches are projects that can develop independently from the same foundation. A converging dependency means the later project follows the combined results of every preceding branch. Projects grouped without arrows share a research topic without a continuation or dependency between them.
Project Families
Recursive vision Transformers
Recursive weight-sharing and token merging offer ways to reduce the parameter and computational demands of vision Transformers. Research examines hardware-aware compression schedules, input-adaptive merging, and spatial reconstruction, assessing accuracy, memory use, and runtime overhead.
Combining Recursive Weight-Sharing with Token Merging for Edge Vision Transformers
Hardware-aware design-space exploration
Input-adaptive token merging
Spatially aware token reconstruction
Transformer deployment on FPGA
This family investigates Transformer inference on FPGA hardware for particle tracking under latency, memory, and accuracy constraints. Research covers model partitioning, reusable hardware blocks, and synthesis workflows, examining communication, buffering, and hardware resource costs.
Automated Transformer Deployment on FPGA for Particle Tracking
Latency improvement for partitioned FPGA deployment
Reusable Transformer block synthesis
Industrial CPS streaming
Continuous monitoring of industrial cyber-physical systems requires timely fault predictions under memory and latency constraints. The research covers stream processing, synthetic data generation, adaptive preprocessing, and specialised model routing, with attention to temporal structure, changing operating conditions, and data quality.
Postmortem, phase-based analysis of industrial CPS time-series data
Runtime stream processing for industrial CPS behaviour monitoring
Generative models for synthetic industrial CPS data streams
Adaptive preprocessing for streaming industrial CPS data
Specialised model routing for streaming fault detection
Information position effects
This family examines how system knowledge and data preparation influence machine-learning predictions for industrial fault and wear detection. Research centres on knowledge-informed segmentation and controlled comparisons that distinguish information effects from data leakage and differences in operating conditions.
Knowledge- and data-position framework for ML-assisted fault identification
Knowledge-informed segmentation for milling tool-wear detection
Model design and efficient distributed deployment
Model structure, data selection, and deployment decisions shape the resource demands of distributed machine learning. This family explores model parallelism, data-efficient architecture search, online data distillation, and frugal inference orchestration, balancing predictive quality and preservation of rare cases with per-node compute, memory, and communication costs.
Transforming Convolutional Neural Networks for Model Parallelism
Neural Architecture Search Optimisations for Data-Efficiency
Automated Data Distillation with Corner Case Preservation
A Research Orchestrator for Frugal Distributed Inference
REDVID simulation framework
REDVID provides configurable, reduced simulations of particle propagation to generate synthetic tracking data. Research explores detector geometry, timing, and event complexity, using reproducible benchmarks to relate simulation detail to computational cost.
REDuced VIrtual Detector simulation framework for particle propagation
Multiple enhancements
Time dimension and 4D tracking
Track generation and complexity levels
Electron simulation
Muon simulation
Performance Analysis and Benchmarking for Simulations
ML deployment across the edge-cloud continuum
This family explores the deployment of serverless workloads across edge and cloud resources. It examines orchestration and workload placement alongside model decomposition for machine-learning inference, with attention to resource utilisation and per-function memory and execution limits.
Towards Enabling Serverless Computing Across the Edge-Cloud Continuum
Designing an Efficient Decomposition Method for the Deployment of Machine Learning Models on Serverless Platforms