Deep Learning Intern · University of Manchester
Prototype
?
Pick two bars — click, or use arrow keys and Enter — to test whether they can “see” each other.
- Intel i5 notebook (serial) ~8–12 h
- CSF4 · 32-core SLURM ~1–2 h
The visibility-graph mapping at the heart of DVS: two bars connect when the line between their tops clears everything in between. The race replays the 10× HPC speedup (time compressed).
Highlights
- Time-Series to Graph: Investigated time-series-to-graph representation learning techniques for long-horizon forecasting tasks to model complex temporal relationships.
- Algorithm Implementation: Developed a custom Deep Visibility Series (DVS) algorithm from scratch using Python, utilizing a Convolutional Neural Network (CNN) backbone for spatial feature extraction on time-series generated visibility graphs.
- Attention Mechanism Tuning: Coded, integrated, and benchmarked multi-head self-attention modules into an LSTM architecture, reducing time-series forecasting model errors by ~3% compared to baseline benchmarks.
- Hybrid Modeling: Engineered a novel hybrid forecasting system combining graph-based DVS encodings with multi-head attention blocks, surpassing customized long-horizon benchmarks by ~2%.
- HPC Parallel Optimization: Ported training workloads from a restricted Intel i5 local notebook (~8–12 hour overnight serial executions) onto the CSF4 Linux HPC Cluster using parallelized SLURM batch configurations (~32 cores), achieving a 10x compute speedup (~1–2 hour runs).
Technologies
- Python
- CNN
- LSTM
- Multi-Head Attention
- SLURM
- CSF4 HPC Cluster