Level 3 · Advanced In Progress¶
Goal: assemble Level 1–2 tools into the math that actually trains neural networks — backpropagation derived from the multivariable chain rule, the derivatives behind every common loss function, momentum/Adam as refinements of vanilla gradient descent, regularization, maximum likelihood estimation as the principle behind most loss functions, softmax/cross-entropy, computational graphs and autodiff, batch normalization, and the numerical stability tricks production ML code relies on.
Modules¶
- Backpropagation from the Chain Rule
- Loss Functions & Their Derivatives
- Optimization Beyond Vanilla GD
- Regularization Math (L1/L2)
- Maximum Likelihood Estimation
- Softmax & Cross-Entropy Derivatives
- Computational Graphs & Autodiff
- Batch Normalization Math
- Numerical Stability in ML Math
- Capstone — Hand-Derive a Small Network
By the end of this level you'll be able to hand-derive backpropagation through a small network, derive gradients for MSE/cross-entropy/softmax, explain why Adam converges faster than vanilla gradient descent, and reason about the numerical pitfalls (overflow, underflow, vanishing/exploding gradients) that separate textbook math from working code — with every derivation checked against NumPy/PyTorch-style autodiff.