About the course
An introduction to NeuroAI, the field that brings brain science and artificial intelligence together. Students build a mathematical and computational toolkit for studying brain, mind and behavior, from classical models to hands-on work with modern machine learning in Python.
The course has three themes: describing representations, from spaces and similarity to geometry and how to read representations out and compare them; learning representations, from the model neuron to deep networks, including vision, model–brain comparison and generative models; and dynamics, sequences and world models, covering recurrence, attractors, attention, language models and world models. Explainability runs through all three. For each model we ask what it computes, how it computes it, and whether the same account holds for the brain.
The course is for advanced undergraduates and graduate students in cognitive science, neuroscience, psychology, computer science and applied mathematics. It assumes some programming experience but no particular disciplinary background.
Slides are posted here the day before each lecture. Each deck also has a text version that works with screen readers.
Lectures
Lecture 0 · Sep 10Welcome & what is NeuroAI
Theme 1Describing representations: spaces, similarity & geometry
Lecture 1 · Sep 15Representations, spaces & metrics
Images, neurons and networks as vectors; noise and averaging; the RDM
Lecture 2 · Sep 17Similarity, MDS & comparing RDMs
Distance measures & the metric axioms; MDS & Shepard's law
Lecture 3 · Sep 22PCA & eigendecomposition
Covariance, principal components, compression
Lecture 4 · Sep 24Nonlinear dimensionality reduction
Manifolds; Isomap, t-SNE & UMAP
Lecture 5 · Sep 29Geometry of neural representations
Dimensionality & linear readouts
Lecture 6 · Oct 1Reading and comparing representations
Shattering, cross-condition generalization, mixed selectivity, RSA & CKA
Lecture 7 · Oct 6Geometry, baselines & tests
Flexible and abstract codes; shuffles, cross-validation, the bootstrap, noise ceilings
Theme 2Learning representations: from neurons to deep networks
Lecture 8 · Oct 8Learning the weights: from neurons to gradient descent
The model neuron, training objectives, the perceptron & gradient descent
Lecture 9 · Oct 13Hidden layers & backpropagation
The chain rule, XOR & the four steps; softmax, autograd & transfer learning
Lecture 10 · Oct 15Training at scale, generalization & characterizing units
Minibatches, schedules & Adam; overfitting & regularization; tuning curves & Gabor fits
Lecture 11 · Oct 20The visual system & CNNs
The ventral stream, V1 and sparse coding, convolutional networks, transfer
Lecture 12 · Oct 27Comparing networks with the brain
Encoding models, Brain-Score and their limits
Lecture 13 · Oct 29Understanding what networks compute
Attribution, feature visualization, shortcut learning
Lecture 14 · Nov 5Autoencoders & generative models
Autoencoders, VAEs, GANs & diffusion
Theme 3Dynamics, sequences & world models
Lecture 15 · Nov 10RNNs, sequences & decisions
Backprop through time; evidence accumulation; LSTMs & GRUs
Lecture 16 · Nov 12Neural dynamics & attractors
Hopfield networks & energy-based models
Lecture 17 · Nov 17Attention & the transformer
Self-attention, transformers & attention maps
Lecture 18 · Nov 19LLMs, self-supervised learning & predictive coding
Next-token prediction & scaling; contrastive and masked learning
Lecture 19 · Nov 24World models & frontiers of NeuroAI
World models; does better AI make better brain models?
Notes
Short notes that back up the lectures and recitations.
- Neuroscience background (PDF)
- Linear algebra (PDF)
- PCA and statistics (PDF)
- The perceptron (PDF)
- Backpropagation (PDF)
- Notation reference (PDF)