Between symbols and vectors.
Neurosymbolic AI is the study of the seam: how discrete structures — logic, graphs, programs, plans — meet continuous learned representations. This course follows the round trip, from foundations to applications.
Most AI courses teach one camp. This one studies the seam.
A neural network can see. A symbolic system can reason. Neither, alone, can do what a robot or a mathematician needs: turn perception into symbols, reason over them, and act on the result. The seam is where the interesting failures — and the interesting ideas — live.
The course is not a tour of techniques. It is one question, followed down five layers: how do discrete symbols and continuous representations meet?
Each answer is a family of methods: logic as a loss, logic as an architecture, logic as data, logic as search, logic as a verifier. Each application is a place where the round trip pays off — or breaks.
We read primary papers, not summaries. Every worked example is tied to a published system and a figure we either redraw or attribute. When the evidence is synthetic, we say so. When a claim is theoretical, we mark it.
The course is self-paced: a lecture is a self-contained artifact, not a deadline. The lab is where the ideas become code.
Discrete, compositional, verifiable. Logic, graphs, programs, plans. Strong where structure matters; brittle where the world is noisy.
Continuous, learned, robust. Embeddings, attention, policies. Strong where perception matters; opaque where guarantees matter.
The hard part. A loss, an architecture, a dataset, a search procedure, a verifier — five different answers to the same question.
Twenty-one lectures, one spine.
Framing → the two worlds → the five interfaces → the three inferences → applications → frontiers. All 21 lectures are live.
Four passes, one idea.
A lecture is not finished until all four exist. Read the argument, move the figure, run the code, then find the gaps in your own understanding.
Read the guide
A long-form argument with the math, the history, and the caveats. Not a summary; a position.
Play with the figure
A browser-native interactive: a scene graph, a program, a constraint solver. Move one variable and watch the system fail or recover.
Run the lab
A notebook or a self-contained browser lab. The Sudoku lab is first: a noisy perception module plus hard rules, solved together.
Recall the gaps
A short quiz with explanations. The score is not the point; the questions you cannot answer are the reading list.
What is neurosymbolic AI?
Lecture 01 is the map: a full formulation, a taxonomy of models, the history and philosophy, the problems the field claims to solve, and seven worked examples — from CLEVR to Sudoku to AlphaGeometry.