Continuous Confidence Labeling & CNN Regression

Welcome to Lesson 33 of the SNAP ADS Learning Hub! This lesson covers the two remaining commitments in the ADACL acronym — Detect Anomalies (the regression framing) and Continuous Labeling (the supervision signal). They sit together because they only work together: a regressor with binary supervision is still a classifier in disguise, and continuous labels with a classifier head are wasted. The combination is what closes the binary-confidence gap we characterised in Lesson 30.

This is the technical core of ADACL’s training story. We will look at three things in order: how a continuous label is computed for a given training example, why a CNN architecture fits the feature stack we feed it, and what the resulting calibrated output means downstream.

A useful image: a thermometer is to a fire alarm as continuous confidence is to binary classification. The thermometer is uninteresting until you ask the question “is the room getting warmer?” — at which point it becomes the only sensible instrument. Anomaly detection is the same: nobody wants a fire alarm; everybody wants the thermometer that tells them they should open a window in twenty minutes.

Step 1: Where Do the Continuous Labels Come From?

The labels are computed, not annotated. For each training example — real or augmented — ADACL evaluates a physics-informed deviation function that maps the example’s full state representation to a scalar in [0, 1]. The deviation function is the bridge between physical reality and the supervised-learning loss.

Three concrete inputs to the deviation function for a typical quantum-system example:

1. The distance from the DeCoN-PINN-predicted baseline. Given the example’s state, what does the trained DeCoN-PINN model say the state should be? The trace distance (or fidelity-derived metric) between observed and predicted is one component of the deviation.
2. The drift parameters applied during augmentation. If the example was synthetically generated by perturbing T₂ by 20 % or adding a 5 % cross-talk term, the magnitude of the perturbation contributes to the label on a calibrated scale.
3. The residual against physical constraints. How much does the example violate conservation, normalisation, or known operating-envelope bounds? Larger residuals get larger labels.

The deviation function is a weighted combination of those components, normalised so that a perfectly-baseline example labels at 0 and a maximally-anomalous-in-distribution example labels at 1. The weights themselves are calibrated against a held-out set of operator-validated examples — operators triage a sample of past alerts as “definitely real anomaly” vs “marginal” vs “false positive,” and the weights are tuned so the deviation function reproduces those judgments.

The crucial property: equal increments in the label correspond to equal increments in physical deviation. A 0.5 example is twice as deviant as a 0.25 example along whatever physical axis matters. This calibration is what makes the eventual output usable for fine-grained policy decisions.

Step 2: Why a CNN, Specifically?

Once we have features and labels, we need an architecture. ADACL uses a convolutional neural network regressor. Three reasons drive that choice over fully-connected or transformer alternatives.

Temporal locality in the features

A large fraction of ADACL’s 394 features (Lesson 37) are computed over rolling time windows of sensor signals. Adjacent feature positions in the input tensor correspond to adjacent time windows of the same sensor. Convolutional kernels naturally exploit this locality: a small kernel sees a few neighbouring time windows and learns local temporal patterns (sudden jumps, gradual drifts, oscillatory regimes). Fully-connected layers would have to learn this locality from scratch and would need orders-of-magnitude more data to do so.

Translation invariance is desirable

A drift event that starts at minute 12 of a window should produce the same anomaly score as the same drift event starting at minute 27 of the same window. Convolutional weights are shared across positions, so the model is translation-invariant by construction. A fully-connected model would have to learn invariance from augmentation alone, which is inefficient.

Parameter efficiency on small datasets

The CNN’s weight sharing dramatically reduces parameter count compared to fully-connected alternatives. For ADACL’s small-label regime, fewer parameters means lower overfitting risk and faster convergence — both critical when labelled positives number in the hundreds.

The architecture is not the heroic ingredient — the labels and the augmentation are — but the CNN is the right match for the rest of the pipeline.

Step 3: The Regression Loss

With continuous labels in hand and a CNN regressor in place, the loss is straightforward: a smooth regression loss between predicted score and label. ADACL uses a hybrid loss with two terms:

• Mean squared error (MSE) on the predicted score against the continuous label — the standard regression objective, well-behaved everywhere in [0, 1].
• Hinge term on the most-anomalous examples — for the top 5 % of training examples by label magnitude, add a hinge penalty if the predicted score is below a fixed margin. This sharpens the model’s response to the genuinely dangerous tail without sacrificing calibration in the middle.

A pure MSE loss would tend to under-predict in the high-anomaly regime (because the squared error is dominated by the abundant low-label examples). The hinge term is the corrective. Together they produce a regressor whose predictions are calibrated everywhere and sharp where it matters.

Crucially, no binary cross-entropy appears anywhere in the loss. The model is never asked to classify; it is only asked to predict the deviation magnitude. This is what keeps the output continuous and calibrated.

Step 4: What the Calibrated Output Unlocks

Once the CNN is trained, ADACL emits scores in [0, 1] whose value is meaningful. The downstream consequences:

• Dashboards can render the score as a continuous gauge or colour gradient. Operators see the trend, not just the alert.
• Paging policies layer thresholds on the score outside the model — a 0.5 might trigger a logging entry, a 0.75 a Slack ping, a 0.9 a page. Adjusting the policy adjusts behaviour without retraining.
• Mitigation strategies can scale their response to the score. Low scores trigger passive observation; high scores trigger active recalibration or workload pausing.
• Trend analysis becomes possible. Operators can ask “is the score creeping up over the past week?” — a question that’s literally unanswerable with binary output.
• Forensic review becomes easier. After an incident, operators can see exactly when the score crossed each threshold, how steeply it climbed, and which features were driving it (the explanation layer in Module 7 connects the dots).

Every one of these downstream capabilities depends on the score being calibrated. That calibration depends on the continuous labels. The continuous labels depend on the physics-informed deviation function. The framework is a chain.

Why This Combination Is the Whole Point

The previous lessons set up the problem — DeCoN-PINN’s prediction doesn’t natively detect anomalies, the binary-confidence detector loses gradient information, augmentation alone doesn’t help if the supervision is still binary. This lesson closes the loop. With continuous labels derived from physics, a regression objective that respects them, and a CNN that learns efficiently on the augmented data, the framework finally delivers a calibrated, continuous anomaly score — the thing the operator actually needs.

If you take one thing away from Module 6, it’s this combination. Augment + Detect Anomalies + Continuous Labeling collapses to a single sentence: a CNN regressor trained on physics-derived continuous deviation labels over augmented examples. That’s ADACL.

Key Takeaways

• Understanding the fundamental concepts: Continuous labels for each training example are computed by a physics-informed deviation function whose output is calibrated to match operator-validated judgments. A CNN regressor then learns to predict that score with a hybrid MSE+hinge loss.
• Practical applications in quantum computing: The score is calibrated such that equal increments correspond to equal physical deviation. A 0.7 score means roughly seven-tenths of the way along the trained deviation scale — actionable, comparable across days, and stable across deployments.
• Connection to the broader SNAP ADS framework: The continuous output unlocks the entire downstream operational layer — dashboards, paging policies, mitigation strategies, trend analysis, forensic review. None of those work without calibrated continuous confidence.

What’s Next?

Lesson 34 walks through the full ADACL training pipeline end-to-end — how the augmentation loop, the labelling function, the CNN regressor, and the feedback layer wire together into a single training and serving system. After that, Module 7 opens the technical deep dive.

---

Ready to continue? Use the navigation buttons below to move to the next lesson or return to the module overview.