6.4 Neural Network Coverage Measures

Statement coverage and branch coverage assume explicitly coded paths: if this condition is true, take that branch. A trained neural network does not read like a nested if tree. Its behavior is dictated by learned weights, bias values, and activations. Two inputs can take “the same route” through the layer list and still light up completely different neurons. Learning objective AI-3.4.3 (K2) asks you to describe the specialized coverage measures that testers use to see how thoroughly test inputs exercise those internal mechanisms.

These measures are structural in spirit — they ask “did this part of the model move?” — but the “parts” are neurons and ranges of activation, not source-code decisions. Typical approaches include the three named below.

Neuron coverage

Neuron coverage is the proportion of neurons in the network whose output exceeds a specified threshold during testing.

Pick a threshold, often a small positive number if you care about “did this unit turn on,” or a task-specific cut if the team already uses one. Run the test set (or a live traffic sample you are allowed to use as tests). For each neuron, record whether its activation ever cleared the threshold. Neuron coverage = (neurons that cleared the threshold at least once) / (neurons you are counting).

If 80 of 100 hidden neurons exceeded 0.5 on at least one test, neuron coverage is 80% at threshold 0.5. The remaining 20 never “fired” by that definition. Those dead or untested units are a testing lead: maybe no test reached the feature they represent; maybe they truly never activate on this data; maybe the threshold was crude. The metric does not tell you which story is true. It tells you where to look and where to invent extra inputs.

Threshold choice matters. A threshold of 0.0 on unbounded ReLU units may count almost everything as covered. A very high threshold may report almost nothing covered even when the net is working. Treat the threshold as part of the test design, documented next to the percentage.

k-multisection neuron coverage (kMNC)

Neuron coverage is a binary “did you ever go above T?” k-multisection neuron coverage (kMNC) is finer. Take each neuron’s possible output range and divide it into k sections. kMNC is the proportion of these sections activated during testing.

Imagine a neuron whose observed (or theoretical) range is −1 to 1, and k = 4. Sections might be [−1, −0.5), [−0.5, 0), [0, 0.5), [0.5, 1]. If every test produces activations only in [0, 0.5), that neuron contributed 1/4 of its sections. Average or aggregate across neurons the way the team’s definition specifies; the syllabus idea you must be able to state is section-level coverage of each neuron’s output range, not just a single on/off bit.

kMNC catches a different gap than neuron coverage. A unit can clear a threshold (neuron coverage satisfied) while only ever sitting in a narrow band of its range. The rest of its dynamic range — strong negative, strong positive — was never exercised. If production later pushes that unit into an unseen section, you have no test evidence for that internal state.

Neuron boundary coverage (NBC)

Neuron boundary coverage (NBC) looks at extremes relative to training, not at a tester-picked threshold. During training, each neuron produced some minimum and some maximum activation. During testing, NBC measures the proportion of neurons whose output either exceeds the maximum achieved in training or is less than the minimum achieved in training.

If hidden unit 17 ranged from 0.10 to 0.80 on the training set, a test activation of 0.95 or −0.02 covers the boundary for that unit. A test activation of 0.40 does not, even if 0.40 is “on” by a neuron-coverage threshold. NBC is asking: did we probe outside the envelope the learner actually saw while fitting?

That question is close to robustness. Production images that are darker, noisier, or simply different can drive activations past training mins and maxes. If NBC stays near zero, your tests may be hugging the training cloud. If NBC is high, tests are leaving that cloud — which is interesting, not automatically “better.” Out-of-envelope activations can be valuable probes or a sign that the test generator is producing nonsense. Testers still have to judge the inputs.

How testers actually use the three numbers

The syllabus use-case is practical:

• Reveal untested neurons or layers. If a whole hidden layer never clears the neuron-coverage threshold, related decision regions may be unprobed. If kMNC shows one section monopolizing a unit, you have not explored that unit’s range. If NBC is zero, you may never have left the training envelope.
• Design additional test inputs or conditions to exercise those underexplored parts. You might add images of a rare lighting condition, messages in a dialect the spam net ignored, or sensor traces at the edge of the operating range — then re-measure coverage to see whether the new tests lit up the dark neurons.

That is test design guided by internals, similar in spirit to covering a previously unhit branch, with the crucial difference that “hit” is an activation statistic, not a line of source.

Tooling is thin, and coverage is not a generalization proof

Commercial tool support for these specific coverage measures is limited. Teams often assemble research code, custom instrumentation, or framework hooks rather than ticking a box in a mainstream GUI. Limited tooling is not an excuse to skip the ideas on the exam; it is a project risk you can name in a test strategy: measuring neuron coverage may require specialist skills.

Structural coverage alone does not guarantee that a neural network will generalize or handle real-world variation. Networks can learn spurious correlations — a watermark, a hospital’s scanner banner, the word “unsubscribe” — and then produce correct-looking activations for incorrect reasons. High neuron coverage can mean “we tickled many units while the model stared at the wrong cue.” Covered and right for the wrong reason is still a defect from a product view.

Therefore testers should combine these measures with other techniques, especially adversarial testing (inputs crafted to fool the model) and metamorphic testing (relations that must hold when you transform an input, even when no single oracle is obvious). Coverage tells you which internal neighborhoods were visited. Adversarial and metamorphic tests stress behavior when the neighborhood is misleading or when a perfect expected label is hard to write.

A short testing story that ties the measures together

You instrument a 3-layer fully connected digit model. Neuron coverage at threshold 0.5 is 72%: twenty-eight percent of neurons never crossed 0.5 on the official test suite. kMNC with k = 10 is 41%: even “live” neurons mostly occupy a few sections. NBC is 8%: almost no test drove a unit past its training min or max. You add darker, zoomed, and slightly rotated digits aimed at the quiet layer. Neuron coverage climbs; NBC ticks up on a handful of units. You still run metamorphic checks (a digit shifted by a few pixels should keep the same class) and a small adversarial set. Coverage moved; that does not close the generalization argument by itself, but you are no longer testing only the neurons that were already loud on clean textbook digits.

AI-3.4.3 in one breath: because nets are weights, biases, and activations rather than coded paths, measure neuron coverage (threshold hits), kMNC (k slices of each output range), and NBC (outside the training min/max). Use the gaps to write more tests. Expect weak commercial support. Never treat a coverage percentage as proof the model learned the right thing — pair it with adversarial and metamorphic testing.

Comparing the three measures without mixing them up

Keep a three-row cheat sheet in your head:

• Neuron coverage — one bit per neuron: did activation ever exceed T on the tests?
• kMNC — k bits of resolution per neuron: which sections of the output range appeared?
• NBC — one bit per neuron of a different kind: did a test activation fall outside the training min/max?

A neuron can be “covered” on the first measure, “narrow” on the second, and “inside the training envelope” on the third, all at once. That is not a contradiction. It is why the syllabus lists different measures rather than a single percentage.

When you write a test strategy for an AI-based classifier, say which measure you will compute, how you will choose T or k, how you will obtain training mins/maxes for NBC without leaking hold-out labels into training (the split rules from 6.1 still apply), and which behavioral techniques will sit beside the structural numbers. Coverage is a flashlight. It is not a certificate that the flashlight was pointed at the right object.