Chapter 4 checkpoint

Ten questions to check the complete neural-network workflow.

Choose one answer per question. Read the feedback for every option, then revisit the relevant episode if the reason is unclear.

Checkpoint

1 · Predictions and targets

Predictions have shape (8, 1) and regression targets have shape (8,). What should you do before subtracting them?

1. A Make their shapes identical, preserving the sample axis.

   Correct. Correct: otherwise broadcasting can produce an 8-by-8 matrix of pairwise differences.

2. B Multiply both by the batch size.

   Not quite. This changes values but does not correct the shape mismatch.

3. C Assume subtraction automatically pairs the rows.

   Not quite. Broadcasting follows shape rules, not the scientific meaning of samples.

Checkpoint

2 · Backward and step

What does loss.backward() do?

1. A Calculate gradients of the loss with respect to tracked parameters.

   Correct. Correct: an optimiser step then uses these gradients to update the parameters.

2. B Choose the next learning rate from the validation set.

   Not quite. That is a separate tuning or scheduling decision.

3. C Update all weights to the optimum.

   Not quite. Backward calculates derivatives; it neither updates parameters nor solves the optimisation problem.

Checkpoint

3 · Gradient storage

Why does an ordinary mini-batch loop clear gradients before a new update?

1. A New gradients otherwise accumulate with previous gradients.

   Correct. Correct: accumulation is useful when deliberate, but differs from a fresh gradient per batch.

2. B The training data must be deleted after every batch.

   Not quite. Clearing gradients does not delete the input dataset.

3. C Clearing gradients resets the learned weights.

   Not quite. The weights are retained; gradient storage is cleared.

Checkpoint

4 · Nonlinearity

Two affine layers are stacked without an activation. What kind of function can their composition represent?

1. A Another affine function.

   Correct. Correct: the matrix products and biases can be combined into one affine transformation.

2. B Every nonlinear function if enough parameters are added.

   Not quite. Width alone does not make a composition of affine maps nonlinear.

3. C Only a constant function.

   Not quite. The output can still vary linearly with the input.

Checkpoint

5 · Binary loss

What should be supplied to BCEWithLogitsLoss for a single binary target?

1. A Raw logits and floating-point zero/one targets with matching shapes.

   Correct. Correct: the loss includes the probability transformation in a stable calculation.

2. B Sigmoid probabilities that have already been rounded to zero or one.

   Not quite. Hard decisions discard information and violate the raw-logit interface.

3. C Integer class indices for a four-class problem and one output.

   Not quite. A mutually exclusive four-class problem usually needs four logits and CrossEntropyLoss.

Checkpoint

6 · Batch averages

Five batches contain 20 samples each and a final batch contains 3. How should you calculate the epoch mean loss from batch means?

1. A Weight each batch mean by its number of samples, then divide by 103.

   Correct. Correct: every observation receives equal weight.

2. B Average all six batch means equally.

   Not quite. This overweights the final three samples.

3. C Keep only the final batch’s loss.

   Not quite. That does not summarise the complete training epoch.

Checkpoint

7 · Evaluation mode

Which pair is appropriate for validation of a network containing dropout?

1. A model.eval() together with torch.no_grad().

   Correct. Correct: one changes module behaviour, the other disables gradient tracking.

2. B Only torch.no_grad(), because it disables dropout.

   Not quite. No-grad does not change dropout’s training behaviour.

3. C model.train() followed by an optimiser step on the validation loss.

   Not quite. That fits the validation data and invalidates its held-out role.

Checkpoint

8 · Early stopping

The selected validation checkpoint is from epoch 30, while patience expires at epoch 42. Which model should you retain?

1. A The independent saved copy of epoch-30 weights.

   Correct. Correct: restore the checkpoint selected by the stated validation rule.

2. B Epoch 42 because it trained longest.

   Not quite. That ignores the selection rule and may retain an overfitted model.

3. C Whichever of the two scores better on the final test set.

   Not quite. This would use the test set to choose an epoch.

Checkpoint

9 · Reproducibility

Which statement about a fixed seed is justified?

1. A It helps repeat a run in the same environment, but does not establish robustness across seeds or environments.

   Correct. Correct: save the procedure and software versions, and assess variability separately.

2. B It guarantees identical results on every future device and library version.

   Not quite. Numerical kernels and implementations can change across environments.

3. C It makes a single test split an exact estimate of future performance.

   Not quite. Fixing randomness does not eliminate sampling uncertainty or distribution shift.

Checkpoint

10 · Choosing the model

A network and ridge regression are close on validation RMSE. What is a defensible next step?

1. A Apply a declared selection rule that considers performance and complexity, then evaluate the fixed procedure once on test data.

   Correct. Correct: a simpler model can be preferable, and the test set estimates the chosen procedure.

2. B Run many network seeds on the test set and report the best one.

   Not quite. This selects on test noise and gives an optimistic estimate.

3. C Choose the network because more parameters guarantee better generalisation.

   Not quite. Capacity is not evidence of usefulness on new observations.

PreviousLab 4 A neural network for solubilityNextChapter 4 key points

Edit this pageReport a problemLast updated 2026-11-04