Selected work from active learners
These projects were completed as part of the neural networks and architecture curriculum. Each one reflects a specific technical problem the student chose to address.
Convolutional Attention for Medical Image Segmentation
A U-Net variant with added channel-wise attention gates, trained on publicly available retinal scan data. The student spent six weeks tuning the loss function weighting before reaching stable training behaviour. The architecture choices are documented with ablation notes.
Transformer Encoder for Source Code Vulnerability Detection
Static analysis reframed as a classification task. The model was pre-trained on a curated corpus of C functions, then fine-tuned on labeled vulnerability data. Positional encoding was adapted to reflect syntactic tree depth rather than token position — an unusual design decision that held up in evaluation.
Conditional GAN for Architectural Floor Plan Synthesis
Given a room count and area constraint as input vectors, the network generates plausible floor plan sketches. Training required careful balancing between the discriminator and generator — the project write-up includes a detailed section on mode collapse and how it was partially resolved through spectral normalisation.
What the numbers reflect
Each metric here comes from project submission records and reviewer notes, not from surveys. The figures describe what students actually produced across all active cohorts since the curriculum reached its current form. Projects are evaluated on architecture clarity, reproducibility of results, and quality of written reasoning — not on benchmark scores alone.
Reviewer comments are returned within eight days. Students may revise and resubmit once before final grading closes.