Top-k Neuron Coverage in Neural Testing

• Top-k Neuron Coverage (TKNC) is a metric that measures the activation of leading neurons in each DNN layer, ensuring test inputs hit the most critical nodes.
• It uses a tunable parameter K to identify high-activation neurons without detailed profiling, enabling efficient evaluation of network behavior.
• Empirical studies show TKNC’s sensitivity to K values, making it a valuable complement to other metrics for safety-critical and robust neural network testing.

Top-k Neuron Coverage (TKNC) is a structural testing metric for deep neural networks that quantifies the extent to which a test suite exercises the “most active” neurons in each layer. Introduced in DeepGauge [Ma et al. 2018], TKNC has gained adoption in structural coverage toolkits such as DNNCov for evaluation, particularly in safety-critical deployments and thorough network testing. The metric captures coverage in terms of high-activation behavior, eschewing per-neuron thresholds or profiling, and serves as a complement to other established coverage metrics. Key variables in its application include the tunable parameter $K$, which governs metric granularity and stringency.

1. Formal Definition and Notation

Let $N$ denote a feed-forward neural network with $L$ layers. For each layer $l$ ($1 \leq l \leq L$), let $n_l$ denote the number of neurons and $a_{l,i}(x)$ be the activation of neuron $i$ in layer $l$ for input $x$. For input $N$0 and integer $N$1, define:

• $N$2

Over a test suite $N$3, neuron $N$4 is “covered” if $N$5 for at least one $N$6.

Denote the set of all covered conditions:

$N$7

The Top-k Neuron Coverage on a test suite $N$8 is:

$N$9

No normalization or thresholding is required beyond the division by the total number of neurons.

2. Coverage Computation and Example

TKNC evaluates coverage by determining, for every neuron in each layer, whether it appears within the top $L$0 highest-activated neurons for at least one test input. Let $L$1 if neuron $L$2 is ever in $L$3 for some $L$4, 0 otherwise:

$L$5

Worked Example:

Consider a network with two hidden layers ($L$6) and no bias. Let $L$7, $L$8, $L$9, and $l$0. Activations:

• $l$1 with $l$2
• $l$3 with $l$4
• $l$5 with $l$6
• $l$7 with $l$8

Coverage:

• Layer 2: $l$9
• Layer 3: $1 \leq l \leq L$0
• $1 \leq l \leq L$1

For $1 \leq l \leq L$2, coverage would be only $1 \leq l \leq L$3 of $1 \leq l \leq L$4 in layer 2 and $1 \leq l \leq L$5 of $1 \leq l \leq L$6 in layer 3: $1 \leq l \leq L$7.

3. Tooling and Implementation: DNNCov Framework

TKNC is implemented in DNNCov, which extends DeepHunter. The main computational steps are:

1. Batch forward-pass the test set, extracting each hidden-layer activation matrix $1 \leq l \leq L$8 (shape: batch$1 \leq l \leq L$9).
2. For each sample and layer, sort the row of $n_l$0 and select top $n_l$1 indices.
3. Maintain a Boolean array $n_l$2, initialized to false; set $n_l$3 if $n_l$4.
4. After all inputs, compute $n_l$5 as $n_l$6.

Key parameters:

• $n_l$7: number of top-activated neurons (user-configurable)
• Layers: by default, every hidden layer from $n_l$8 to $n_l$9
• Batching and vectorized operations enable 2.5$a_{l,i}(x)$0 speed-up compared to sequential calculation

No training-set profiling or per-neuron thresholds are required, in contrast to KMNC, NBC, or SNAC.

4. Empirical Evaluation and K Sensitivity

Empirical assessment in (Usman et al., 2022) demonstrates the impact of $a_{l,i}(x)$1 on TKNC for LeNet-1, LeNet-4, LeNet-5, ResNet20, and TinyTaxiNet. Results are summarized as follows:

Model	TKNC ($a_{l,i}(x)$2)	TKNC ($a_{l,i}(x)$3)
LeNet-1	88.57%	1.00%
LeNet-4	81.59%	3.27%
LeNet-5	82.40%	4.93%
ResNet20	65.09%	3.90%
TinyTaxiNet	52.06%	0.59%

With a small $a_{l,i}(x)$4, the majority of neurons are covered, yielding high coverage (50–90%). For large $a_{l,i}(x)$5, coverage collapses to a few percent, even on test suites of reasonable size. This demonstrates TKNC's pronounced sensitivity to the choice of $a_{l,i}(x)$6 and the need for careful parameterization.

5. Selecting K and Relation to Other Metrics

Best practices for $a_{l,i}(x)$7 selection include:

• Avoiding trivially small $a_{l,i}(x)$8 (e.g., $a_{l,i}(x)$9), which rapidly saturates coverage
• Avoiding excessively large $i$0, which may exceed layer widths and result in zero coverage for some layers
• In convolutional networks, choosing $i$1 as a small constant (5–20), or as a fixed proportion of $i$2 (e.g., top 10%)
• Using validation data to cross-validate $i$3 so that TKNC falls in an informative regime (recommendation: 30–70% coverage)

Comparison to related structural metrics:

• NC (Neuron Coverage): counts neurons with activation $i$4 at least once. Easily saturated; no activation ranking.
• KMNC (K-Multisection Neuron Coverage): divides the profile range of each neuron into $i$5 bins, measuring finer granularity but requiring per-neuron bound profiling.
• NBC/SNAC: focus on boundary conditions, checking for activation past training set minima/maxima.
• TKNC: exclusively targets high-activation neurons, does not require profiling, and omits low or moderate activation cases.

Combining coverage metrics can offer a more complete internal test adequacy assessment.

6. Limitations, Applications, and Recommendations

TKNC foregrounds highly responsive (“hot-spot”) neurons, ensuring tests exercise regions of maximal activation, but is insensitive to boundary or moderate activations. Thus:

• It is optimizing for high-activation scenarios, underrepresenting rare or subtle neuron behaviors.
• Does not capture low-activation or edge-case behaviors, unlike NBC/SNAC.
• For comprehensive coverage—especially in safety-critical systems—joint use with value-range section coverage (KMNC) and boundary-focused metrics (NBC, SNAC) is recommended.
• When interpreting test adequacy, especially in functional safety contexts, TKNC should be correlated with cause-effect reasoning coverage such as MC/DC variants.

TKNC is a lightweight, parameter-free (other than $i$6) metric, designed for efficient, practical measurement of neural test coverage in modern DNNs, with clear empirical behavior and tooling support in DNNCov (Usman et al., 2022).