Skip to content
ML Academy

Week 7 · Sunday — consolidate and defend

Neural project: explain an update, read the curves, change one thing

Project thread neural · linked lessons day 31, day 32, day 33, day 34, day 35

Milestone N2 of the neural thread: a diagnosed training curve, one revised training choice re-run on the frozen split, and an explained single-weight update.

  • 90 minsession length
  • 4timed blocks
  • 5tasks
  • —status

Session plan90 min

  1. 20′Spaced recall
  2. 40′Project improvement
  3. 20′Teach-back and decision defense
  4. 10′Weekly review and planning

Before you start

Before the session: Saturday's project repository with data/split_v1.npz, the training script that reproduces Saturday's run from the command line in under two minutes on a laptop CPU, and the saved loss log. No new data or installation. If Saturday's run is unfinished, the setup is that unfinished script; see carry-forward.

Blocks4

  1. Block 1: Spaced recall

    20 min

    Closed notes, then check and correct in a second colour:

    • Days 31-32: parameter count for 64 -> 32 -> 10; which output layer and loss for multi-class, binary and regression targets; why logits are passed to a numerically stable softmax cross-entropy.
    • Day 33: one sentence each separating backpropagation (computing gradients) from the optimizer (using them).
    • Day 34: the curve pattern for divergence, underfitting and overfitting, with one intervention each.
    • One older item: where in your Saturday code the scaler was fitted, and why fitting it on all rows would be leakage even for an image dataset.

    Log each miss with the misconception it reflects; these feed your review queue.

  2. Block 2: Project improvement

    40 min
    1. Diagnose (10 min). Read your Saturday train/validation loss log. Classify it as diverging, underfitting, overfitting, or healthy, and cite the epochs that support the label. If the curve looks healthy, create a diagnosis case on purpose: re-run once with a learning rate about ten times larger and diagnose that curve.
    2. Pre-register (5 min). Choose one training choice to revise (learning rate, number of epochs with early stopping on validation loss, weight decay, or hidden size). Write the expected effect and the result that would prove you wrong before running.
    3. Run (20 min). Change only that one choice, keep the seed, split and metric fixed, and re-run. If time allows, repeat the old and new configuration over three seeds to see whether the difference exceeds seed-to-seed variation.
    4. Record (5 min). Add a row to runs/experiments.csv: configuration, seed(s), validation metrics, and the verdict against your pre-registered expectation.

  3. Block 3: Teach-back and decision defense

    20 min

    Pick one weight in the output layer. Using values printed from your own network for a single training example, explain aloud (record two to three minutes, or write 150 words): the forward value, the local gradient from the softmax cross-entropy output (prediction minus one-hot target), the chain-rule step to that weight, and the update applied by your optimizer. Then defend your revised training choice in three sentences: what the curve showed, what you changed, and what the evidence says now.

  4. Block 4: Weekly review and planning

    10 min

    Update your weekly log: corrected recall misses, the experiment verdict, unfinished items with their studio ID, and next week's priority. For Week 8 the priority is fixed in kind (one controlled regularization or learning-rate experiment on the same split) but you choose which factor based on today's evidence.

Tasks5

Sign in to save your work.

  1. Task 1

    recall

    Answer the closed-notes recall set (parameter count, output layer and loss pairing, backpropagation vs optimizer, curve patterns, scaler leakage), then correct each miss.

    The expected evidence for this task is shown after you save your own output.

  2. Task 2

    analysis

    Classify the Saturday loss curve (or a deliberately high-learning-rate run) as diverging, underfitting, overfitting or healthy and cite specific epochs and values from the log.

    The expected evidence for this task is shown after you save your own output.

  3. Task 3

    code

    Pre-register a single training change, re-run with seed, split and metric fixed, and compare old and new on validation; add multiple seeds if time allows.

    The expected evidence for this task is shown after you save your own output.

  4. Task 4

    defense

    Explain one output-layer weight update end to end with numbers from your own network and one example.

    The expected evidence for this task is shown after you save your own output.

  5. Task 5

    writing

    Update the weekly log and set the Week 8 priority.

    The expected evidence for this task is shown after you save your own output.

Rubric4 criteria

CriterionMeetsExceeds
Curve diagnosisThe label (diverging, underfitting, overfitting or healthy) is supported by cited epochs and is consistent with both train and validation curves.Names and rules out a plausible rival explanation, such as a data or preprocessing bug, with a specific check.
Controlled changeExactly one training choice changes; seed, split and metric stay fixed; the expectation is written before the run; the verdict follows from the numbers.Uses several seeds and reports the spread, and says honestly when the difference is within that spread.
Update explanationThe update is computed with the learner's own values and correctly separates the gradient computation from the optimizer rule.Explains how the same gradient would produce a different step under momentum or Adam.
Review and planningMisses are corrected and tagged, unfinished items are carried forward with IDs, and the Week 8 priority is specific.The priority is tied to a stated hypothesis that Week 8 Saturday can confirm or refute.

Required output

Corrected recall mistakes, an improved or explicitly unimproved training configuration with its pre-registered verdict, a short evidence-based explanation of one weight update, and next week's priority. Together with Saturday this completes the Week 7 evidence: executed small network, shape/update checks, and an explained training result.

Studios are tracked separately from the 60 core lessons; completing one never completes a lesson.