/ ML, DL & GenAI / Phase 4 Phase 4 Days 65-84
Neural Networks From Scratch to PyTorch Micrograd, manual backprop, MLPs, initialization, optimizers, normalization, PyTorch, debugging, tracking, scaling
Phase Goal Close the neural-network intuition gap: build autograd and backprop from scratch, then learn PyTorch training loops, debugging, tracking, and scaling basics.
Day 65: Perceptrons and Computational Graphs
Evaluation & Failure Modes Gradient sanity. Why the forward pass feels easy and the backward pass quietly destroys most beginners. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 66: Micrograd: Scalar Autograd From Scratch
Evaluation & Failure Modes Autograd correctness: testing against finite differences and matching PyTorch element by element. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 67: Backpropagation by Hand
Evaluation & Failure Modes Gradient checks as a permanent debugging tool, not a one-time exercise. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 68: Vectorized MLP From Scratch
Evaluation & Failure Modes Learning dynamics: the four shapes a loss curve takes on MNIST and what each one means. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 69: Activation Functions and Signal Flow
Evaluation & Failure Modes Signal health: when activations explode, vanish, or collapse to a single value. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 70: Initialization Theory
Evaluation & Failure Modes Training stability and why "it diverges at epoch 0" is almost always an init bug. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 71: Loss Functions for Deep Learning
Evaluation & Failure Modes Loss bugs: the silent NaN, the silent inf, and the gradient that is technically correct but useless. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 72: SGD, Momentum, RMSProp, Adam
Evaluation & Failure Modes Optimization curves and the three early warning signs of an optimizer fighting your loss. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 73: Normalization and Regularization
Evaluation & Failure Modes Regularization behaviour: when dropout helps, when it hurts, and the rules of thumb for combining it with batchnorm. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 74: PyTorch Tensors and Autograd
Evaluation & Failure Modes Tensor failure modes: silent device mismatches, in-place ops breaking autograd, and the detach you forgot. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 75: nn.Module, Optimizers, Training Loops
Evaluation & Failure Modes Loop correctness: the seven things that go wrong in a hand-written training loop and how to test for them. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 76: Datasets, DataLoaders, Batching
Evaluation & Failure Modes Data bottlenecks: how to spot them with the torch profiler in five minutes. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 77: Deep Learning Debugging
Evaluation & Failure Modes Training failures: the symptom -> cause -> experiment table you should never have to derive twice. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 78: Experiment Tracking and Configs
Evaluation & Failure Modes Run comparison: the report format that makes experiments legible to a future you. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 79: MLPs for Tabular and Embeddings
Evaluation & Failure Modes Tabular DL tradeoffs: when neural nets actually beat boosters and the data shapes where they cannot. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 80: Keras and High-Level APIs
Evaluation & Failure Modes Abstraction limits: the four moments you have to leave Keras for raw PyTorch. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 81: GPU, Mixed Precision, Memory
Evaluation & Failure Modes Compute reliability: NaN/inf in fp16, loss scaling, and the bf16 escape route. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 82: Distributed Training Awareness
Evaluation & Failure Modes Scaling assumptions that quietly break: batchnorm across processes, learning-rate scaling, and seed sync. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 83: Deep Learning Methodology
Evaluation & Failure Modes Methodology quality: the report that wins a code review. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 84: Capstone: Neural Network From Scratch to PyTorch
Evaluation & Failure Modes Neural-net project quality: the structure that survives a senior review. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Capstone: Neural Network From Scratch to PyTorch Build a tiny neural net from scratch, prove gradients, then rebuild it in PyTorch and train a real classifier with debugging notes and error analysis.
After this phase, you'll be able to:
Implement scalar autograd and manual backprop Train vectorized MLPs from scratch and in PyTorch Debug neural network training failures Track experiments and explain learning curves You can reason about neural nets internally instead of treating frameworks as magic. Gate: explain without notes, build one independent artifact, diagnose a deliberate failure, and repeat a changed task after a delay. Record help and repair missing prerequisites.