Arc 04 · AI Engineer
Neural Systems from Computation Graphs
Build neural networks as computation graphs with local derivative contracts, then diagnose optimization and representation failures.
Exit capability: Implement, test, and debug a neural training loop from first principles.
- Mapped lessons
- 7
- Published now
- 1
- Full-arc estimate
- ≈16 hours
- Last edited
- 2026-08-11
Lesson sequence
Live units open into complete labs. Planned units stay visible to show the dependency path, but intentionally have no detail route.
- 01
A Neuron Is a Parameterized Decision Surface
PlannedConnect affine maps and nonlinearities to representation capacity.
Concept labFoundation80 min estimate
- 02
Computation Graphs Make Learning Inspectable
PlannedRepresent forward values, dependencies, and local derivatives explicitly.
Build labIntermediate90 min estimate
- 03
Backpropagation as Local Contracts
LiveDerive reverse-mode autodiff as small, testable vector-Jacobian products rather than one mysterious global formula.
Build labIntermediate120 min estimateArtifact: Tiny autodiff engine and tests
- 04
Read Optimization Dynamics
PlannedUse loss curves, gradient statistics, and parameter updates to separate data, optimization, and capacity failures.
Failure labIntermediate105 min estimate
- 05
Normalization and Residual Paths
PlannedUnderstand why signal and gradient paths shape trainability in deep networks.
Concept labAdvanced105 min estimate
- 06
Convolutions and Locality as an Inductive Bias
PlannedBuild a convolution and connect weight sharing to images, audio, and structured grids.
Build labIntermediate105 min estimate
- 07
Neural Training Incident: Loss Becomes NaN
PlannedTrace a realistic failure through inputs, precision, activations, gradients, and optimizer state.
Failure labAdvanced110 min estimateArtifact: Training incident report