Deep learning is the most widely shared method in the program — six scholars use it, on problems ranging from connectome classification to spectrogram regression. This page is the entry point; the more specific approaches have their own pages under Research Methods.
Neuroimaging poses a particular difficulty for these models: sample sizes are small by machine learning standards (hundreds to thousands of subjects, not millions), the data are high-dimensional, and clinical labels are often imbalanced. Most of the strategies below are responses to that constraint — reusing representations learned elsewhere, or choosing an architecture that respects the structure of the data instead of flattening it.
CNS Projects¶
Josh Neudorf — graph neural networks on structural and functional connectomes
Hao-Ting Wang — transfer learning from pooled datasets into a smaller clinical cohort
Tom George — self-supervised pretraining of a neuro-foundation model
Mohamed Abdelhack — modelling visual inference and how perturbations emulate psychiatric conditions
Vibujithan Vigneshwaran — causal deep learning for neuroimaging, with a focus on interpretability and generalization
Lindsay Munroe — explainable AI for predicting neurodegenerative disease progression
Specific approaches¶
Graph neural networks — when the data is a network
Foundation models and transformers — pretrain broadly, fine-tune narrowly
Transfer learning — borrowing representations when your clinical sample is small
Normative modelling — a statistical alternative worth knowing about before reaching for a network
Geometry-aware methods — respecting the curved space that connectomes actually live in
Software and Tools¶
PyTorch Geometric — graph neural networks, with a worked neuroimaging example
BrainLM — a brain language model for fMRI time series
Code environments — reproducible dependency management, which matters more than usual once CUDA is involved
Containers — pinning an entire software stack
Compute Canada / DRAC and HPC jobs — where the GPUs are
Working with imbalanced data — a practical discussion of the class-imbalance problem common to clinical prediction