Training Dynamics & Regularisation Clinic

P6.dl-architectures.10 · Audience: guest, language-pro, it-ml · Prerequisites: Regularisation and Normalisation

Real LLM grading for this pageLLM grading (this page):

Everything in this track taught you a mechanism — backprop, SGD, Adam, dropout. The clinic teaches the craft: reading a run's behaviour and deciding what to change. This is the capstone of Deep-Learning Architectures, and it ends with the platform's first real PyTorch training rungs, graded in the deep-learning kernel.

Step 1 / 3Reading a loss curve
ⓘ Concept: The four silhouettes every practitioner knows
  • Healthy: train and validation fall together, then flatten. Ship it.
  • Overfitting: train keeps falling, validation turns upward — the module-05 story. Regularise or stop early.
  • Diverging: the loss jumps up and pins at a ceiling, or explodes to NaN. Almost always the learning rate.
  • Stuck: the loss flatlines high from step one — a dead architecture, a zero learning rate, or a bad initialisation (a local minimum).

Why it matters — A loss curve is the run's vital sign. Recognising its silhouette in the first seconds tells you what to try next — before you waste an hour of compute.

Replay the overfitting silhouette from module 03 — move the overfit-start slider and watch the validation curve turn:

Ask the mentor about this module

Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.

Ctrl/Cmd + Enter to send

🎓 Practice ladder

3 graded rungs · ~37 min

The capstone ladder. Rung 1 runs in your browser; rungs 2 and 3 are the platform's first PyTorch rungs — your code executes server-side in the sandboxed deep-learning kernel (CPU torch, metered, no network).

Rung 1 — One SGD step, one Adam step (by hand)

Loading exercise…

Rung 2 — Train a tiny torch MLP on XOR (kernel)

Loading exercise…

Rung 3 — Diagnose the diverging run (kernel)

Loading exercise…

Try it yourself

A scratch console for this page's ideas — ungraded, nothing you run here is recorded.

Scratch console

A scratch console with the scientific stack (pandas, numpy, scikit-learn). Runs on the server — no network, resource-limited and measured.

Output appears here.