Softmax Confidence

The softmax function converts a vector of raw logits into a probability distribution over all possible modulation classes. Each output value is bounded between 0 and 1, and the sum of all outputs equals exactly 1.0. This normalization allows direct interpretation: a score of 0.92 for QPSK means the model assigns 92% probability to that hypothesis. The mathematical formulation is σ(z)_i = e^(z_i) / Σ e^(z_j), where the exponential function ensures all outputs remain strictly positive and amplifies differences between the largest logit and its competitors.

The highest softmax value serves as the model's reported confidence in its primary prediction. In modulation classification, this single scalar is often used for downstream decision logic:

• High confidence (>0.95): The classifier is certain; the result can be trusted for automated reconfiguration
• Moderate confidence (0.6-0.95): The prediction is tentative; may warrant additional verification
• Low confidence (<0.6): The signal is ambiguous; the system should request more samples or flag for human review This thresholding enables cognitive radios to adapt their behavior based on classification reliability.

A temperature parameter T can be applied to the softmax function to control the sharpness of the output distribution. The modified formula becomes σ(z)_i = e^(z_i/T) / Σ e^(z_j/T). Key behaviors:

• T < 1.0: Sharpens the distribution, making the model appear more confident by amplifying differences between logits
• T = 1.0: Standard softmax behavior
• T > 1.0: Softens the distribution, producing more uniform probabilities In modulation recognition, temperature scaling is often tuned during model calibration to align reported confidence with actual empirical accuracy, correcting for overconfident predictions on noisy or fading channels.

Raw softmax outputs are often poorly calibrated, meaning a reported confidence of 0.90 does not correspond to a 90% empirical accuracy. This miscalibration is dangerous in electronic warfare and spectrum management applications where decisions carry high consequence. Expected Calibration Error (ECE) quantifies this gap by binning predictions by confidence and measuring the difference between average confidence and observed accuracy. Post-hoc calibration techniques such as Platt scaling or isotonic regression learn a mapping from raw softmax scores to well-calibrated probabilities using a held-out validation set, ensuring that when the classifier says it is 95% certain, it is correct 95% of the time.

Beyond the maximum probability, the entropy of the full softmax distribution provides a richer measure of classification uncertainty. Shannon entropy is calculated as H = -Σ p_i * log(p_i). A peaked distribution (one class near 1.0) yields low entropy, indicating decisive classification. A flat distribution across multiple classes yields high entropy, signaling genuine ambiguity—perhaps due to low SNR or a modulation scheme at the boundary between two known types. In real-time spectrum monitoring systems, entropy thresholds can trigger adaptive sample collection: if entropy exceeds a configured limit, the system captures additional IQ samples to resolve the ambiguity before committing to a classification decision.

In open set recognition scenarios, the classifier must identify when an input belongs to an unknown modulation type not seen during training. Softmax confidence provides a natural rejection mechanism: if the maximum probability falls below a calibrated threshold, the system rejects the sample as out-of-distribution. However, softmax scores for unknown classes can be misleadingly high due to the closed-world assumption baked into the function. More robust approaches combine softmax confidence with energy-based scores or distance metrics in feature space to detect novelty. A common deployment pattern uses dual thresholds: softmax confidence for known-class discrimination and a separate anomaly score for open-set rejection.

A post-processing calibration technique that divides the logits by a scalar temperature parameter before applying the Softmax function. A temperature T > 1 softens the probability distribution, reducing overconfidence, while T < 1 sharpens it. This is critical for calibrating a classifier's expected calibration error (ECE) without changing the argmax prediction.

• Does not alter accuracy, only confidence scores
• Optimal T is found by minimizing NLL on a validation set
• Essential for reliable open-set rejection thresholds

The raw, unnormalized scores output by the final fully connected layer of a neural network before the Softmax activation. Each element corresponds to one modulation class. The logit magnitude reflects the model's uncalibrated evidence for that class.

• Directly manipulated by adversarial attacks to fool classifiers
• Used in knowledge distillation as soft targets
• Logit differences, not absolute values, determine Softmax output

The standard training objective that measures the dissimilarity between the predicted Softmax distribution and the one-hot true label. Minimizing cross-entropy is equivalent to maximizing the log-likelihood of the correct class.

• Loss = -log(p_true_class) where p is the Softmax output
• Heavily penalizes high-confidence errors
• Can be combined with label smoothing to prevent overconfidence

An extension of Softmax designed for open-set recognition, where unknown modulation types may appear at inference. OpenMax recalibrates the Softmax scores using distances to class mean activation vectors (MAVs) and adds an explicit 'unknown' probability.

• Fits a Weibull distribution to extreme value statistics per class
• Rejects inputs far from any known class centroid
• Critical for spectrum monitoring where novel signals are common

A Bayesian approximation technique that runs multiple stochastic forward passes with dropout enabled at inference time to produce a distribution of Softmax outputs. The variance across passes quantifies model uncertainty, while the mean provides the prediction.

• Epistemic uncertainty: model uncertainty due to lack of knowledge
• Aleatoric uncertainty: inherent noise in the IQ data
• Computationally more expensive than a single Softmax pass

An alternative to Softmax confidence that uses the Helmholtz free energy of a model's logits. Lower energy indicates higher confidence and better out-of-distribution detection. Energy scores often outperform maximum Softmax probability for detecting anomalous signals.

• E(x) = -T * log(sum(exp(logit_i / T)))
• Does not require access to outlier data for tuning
• More aligned with the model's density estimation than Softmax