Courses
18 / 18Machine learning: learning a function from examples
What a model is, and how learning a function from examples differs from writing one.
Features: turning raw data into useful numbers
How raw data becomes useful numbers, from word counts to edge filters.
Neural networks: functions that can bend
Build a small network by hand and see how ReLU units let it bend.
Loss functions: measuring how wrong a model is
Turn predictions into one number: mean squared error, softmax and cross-entropy.
Training models: walking down the loss map
Gradient descent, learning rates, initialization and backpropagation.
Evaluating models: will it work on new data?
Training, validation and test sets, overfitting, and common metrics.
Deep networks: why depth helps
Why stacking layers multiplies linear regions, and what depth does and does not buy.
Backpropagation: every gradient in one backward pass
The chain rule in matrix form, its cost in operations and memory, and what it says about initialization.
Regularization: fitting less to generalize better
Weight decay, early stopping, ensembles, dropout, augmentation and double descent, on one running example.
Convolutional networks: learning the filters
Locality and weight sharing, channels, receptive fields, a complete classifier, and a kernel learning to become a Sobel filter.
Residual networks and normalization: training very deep networks
Why plain deep networks degrade, how skip connections and normalization fix it, and what each does to the signal.
Transformers: attention over a sequence
Self-attention worked by hand, scaling, positions, multi-head attention, and a full count of GPT-2.
Agents and workflows: from one model call to a system
What makes a system an agent, and three workflow patterns built in plain Python: prompt chaining, routing and parallelization.
Tool use: letting the model act
How a model asks for tools and a program runs them, the agent loop, ReAct, MCP, and why fetched text can hijack an agent.
Reflection: checking and revising the model's work
A producer-critic loop on the running example, why a critic needs evidence, and what Self-Refine, Reflexion and CRITIC show.
Planning: deciding the steps before taking them
Plan-and-execute with checkable goals, retries and replanning, reasoning before acting, and when to hand planning to a solver.
Memory and retrieval: what the agent knows, and when
Short-term context, retrieval-augmented generation computed by hand, scored long-term memories, MemGPT and skill libraries.
Multi-agent collaboration: teams of specialized agents
A supervisor, parallel researchers, a checker and a human approval gate, plus topologies, debate, A2A and how agents are evaluated.