Entry and exit checks

Español · Facilitator guide · Teach this workshop

Answer alone. Short sentences are enough. No code needed. The learner prompts are also in Notebooks 00 and 12.

Entry · 3 minutes

X.shape == (20, 8, 8).

1. Give two possible meanings for its axes.
2. Predict X[:, 3, 4].shape. Name the remaining axis in each interpretation.
3. What evidence would distinguish independent images from video frames?

Exit · 5 minutes

X.shape == (6, 4, 3, 8, 8): plants, visits, channels, rows, columns.

Spend about three minutes on 1–3 and two minutes on 4.

1. Predict X[:, -1, 1].shape. Name its axes.
2. Predict X.mean(axis=1).shape. Name its axes.
3. What could averaging visits hide? Give one test for that loss.
4. A Tucker model compresses the visits mode from length 4 to rank 2. After reconstruction, does the visits axis have length 2 or 4? Explain why restoring the shape does not guarantee that every brief disease signal survives.

Facilitator key and rubric · reveal after collection

Entry: (images, rows, columns) or (time, rows, columns) both fit. The selection has shape (20,): one pixel across images or time. Acquisition records, timestamps and sample IDs establish the meaning; shape alone cannot.

Exit: selection (6, 8, 8) means plants, rows, columns: last visit, channel 1. The mean has shape (6, 3, 8, 8): plants, channels, rows, columns. Averaging can hide a brief disease signal or growth trend. Compare the visit sequence with its average, using a known transient or a downstream detection test. Other concrete, defensible tests are valid.

Transfer (question 4): the reconstructed visits axis has length 4. The visits factor has shape (4, 2) and expands the core’s length-2 mode back to four visits. Compression can discard variation outside the retained subspace, including a brief disease signal. Restoring shape does not restore every value.

Score each dimension 0 (missing/wrong), 1 (partly right), or 2 (correct and explained). Score the three dimensions below separately for entry and exit, each out of 6. Record question 4 as a separate transfer score out of 2; it has no entry counterpart.

Dimension	Entry evidence	Exit evidence
Axis meaning	Both interpretations, all axes named	Remaining axes named in both results
Shape reasoning	Correct selection shape and why	Both output shapes and why
Evidence	Metadata that establishes batch versus time	A specific loss and a test that can detect it

If axis or shape reasoning scores 0–1, revisit the slice in Notebook 01. If evidence scores 0–1, use group task 02.

For transfer, award 0 for a wrong/missing axis length, 1 for length 4 with an incomplete explanation, and 2 for length 4 with both the factor’s expansion and possible loss of signal explained. For a score of 0–1, revisit core and factor shapes in Notebook 10. If all four dimensions score 2, assign a take-home chosen by interest. Do not treat score differences as a causal measure of workshop impact.

Checkpoints during the lesson

Use the final minute of the core activities in Notebooks 03 and 06 for these individual, closed-solution checks. They replace part of the activity’s explain/check time; they do not add questions or time to the five-minute exit. Collect the first attempt before discussion or revealing the key.

Notebook	Independent prompt
03 · Broadcasting	X.shape == (3, 2) and per-sample offsets b.shape == (3,). Does X - b work? Write the corrected expression and output shape. Explain which axis repeats.
06 · Contraction	S.shape == (2, 3, 4) means batch, time, feature; w.shape == (4,). Write an einsum that retains batch and time. Give the shape and contracted index. What changes with 'ntf,f->n'?

Checkpoint key and next teaching step · reveal after collection

Check	Evidence for 2 points	If incomplete, revisit
Broadcasting	X - b fails: trailing lengths 2 and 3 conflict. X - b[:, None] has shape (3, 2); one offset per sample repeats across feature columns.	Align (3, 2) and (3, 1) on paper, label sample and feature, then retry with different dimensions.
Contraction	np.einsum('ntf,f->nt', S, w) gives (2, 3) and sums f. 'ntf,f->n' also sums time t, giving (2,).	Cross out indices absent from the output, name the information each reduction removes, then retry a new output signature.

Record each checkpoint separately: 0 = missing or wrong operation; 1 = correct operation with an incomplete shape or axis explanation; 2 = all evidence in the table. Use 0–1 to choose the follow-up, then ask for another independent attempt. Keep these formative scores separate from the entry/exit comparison.