stage 05 · Core
Deep Learning
Own neural mechanisms, PyTorch execution, training diagnosis, major architectures, representation learning and reliable measurement on compact models.
8 sections 60 units about 244 h 0/60
Sections
In dependency order
- 1 Neural Computation, Losses & Reverse Mode 0/7
- 2 PyTorch Foundations & Low-Level Training Loops 0/8
- 3 Training Dynamics, Regularization & Diagnosis 0/8
- 4 Convolutional & Recurrent Architectures 0/7
- 5 Representation, Self-Supervision & Paired Modalities 0/7
- 6 Generative Modeling 0/7
- 7 Attention & Transformer Foundations 0/7
- 8 Reliable PyTorch Execution, State & Profiling 0/9
Capstone
Neural Mechanisms & Training Evidence Pack (10 to 16 h)
Review the stage through three compact artifacts that prove mechanism fluency, PyTorch ownership and diagnosis. Keep every model and dataset small enough to rerun locally or in a free notebook runtime.
Milestone deliverables
- Neural core proof pack - branched tensor autodiff with tests, stable cross-entropy, one tensor MLP, finite-difference and PyTorch differential checks, and one saved-value or broadcast defect with its regression test
- Reproducible PyTorch training run - one small CNN, recurrent, VAE or attention model with explicit loops, configuration and split provenance, tiny-set and shuffled-label controls, complete checkpoint state and a stated resume-equivalence result
- Diagnosis and profile casebook - three seeded-fault postmortems across objective or data, numerical or state, and performance or compile behavior, each with hypotheses, a discriminating control, repair, regression guard, measured evidence and one limitation
Rubric
| Criterion | Pass | Strong |
|---|---|---|
| Neural core proof pack | The section 1 mini-project meets Pass on every criterion. | It meets Strong on gradient correctness and on saved values, and one command runs every differential check. |
| Reproducible PyTorch training run | One small CNN, recurrent, VAE or attention model trains in the section 2 trainer with config and split provenance, tiny-set and shuffled-label controls, complete checkpoint state and a stated resume-equivalence result. | Resume is exact on CPU, and a fresh clone reproduces the reported metrics within the stated tolerance. |
| Diagnosis and profile casebook | Three seeded-fault postmortems, one each for objective or data, numerical or state, and performance or compile behavior, each with hypotheses, a discriminating control, repair, regression guard, measured evidence and one limitation. | Each regression guard fails on its fault and passes on its repair in one test command. |
| Rebuild from a clean clone | A README lists the commands that rebuild each artifact. | One script rebuilds every reported number on a CPU or a free notebook GPU, with its run time stated. |
| Honest limits | Each artifact states what its compact model and data cannot show. | Each limit names the later stage or track that addresses it. |
Capstone hours · 10 to 16
Begin the first section