Epistemic Uncertainty

Epistemic uncertainty is fundamentally reducible with more data. Unlike aleatoric uncertainty, which stems from inherent noise, this uncertainty shrinks as the model observes more examples in sparse regions of the feature space. In modulation recognition, a classifier trained on few high-SNR QAM64 samples will exhibit high epistemic uncertainty when encountering a new QAM64 signal with slight channel distortion—a gap that closes with additional diverse training samples.

High epistemic uncertainty serves as a primary trigger for novelty detection. When a model encounters a modulation scheme absent from its training set, the posterior distribution over network weights produces divergent, high-variance predictions. This predictive disagreement—quantified through mutual information or ensemble variance—reliably flags unknown classes like emerging 5G waveforms without requiring prior exposure to them.

Quantifying epistemic uncertainty requires approximating the posterior distribution over model parameters. Key techniques include:

• Monte Carlo Dropout: Applying dropout at inference to sample multiple predictive distributions
• Deep Ensembles: Training multiple networks with different initializations and tracking prediction variance
• Bayesian Neural Networks: Placing explicit priors over weights and learning posterior distributions Each method trades computational cost for uncertainty estimation fidelity.

Epistemic uncertainty concentrates in regions of the embedding space far from training data. In a modulation classifier's penultimate layer, known classes form dense clusters. Inputs mapping to interstitial or distant regions—such as a novel hybrid modulation scheme—trigger high uncertainty because the model lacks support to interpolate confidently. This geometric interpretation enables distance-based rejection using Mahalanobis metrics or prototype networks.

Epistemic uncertainty can be isolated by decomposing predictive entropy: Total Uncertainty = Aleatoric + Epistemic. The mutual information between model parameters and predictions—computed as the difference between entropy of the average prediction and average entropy across posterior samples—isolates the epistemic component. This clean separation prevents confusing inherent signal noise with genuine model ignorance.

Epistemic uncertainty directly guides data acquisition strategies. In spectrum monitoring systems, samples with the highest epistemic uncertainty are prime candidates for human labeling, as they represent regions where the model's knowledge is weakest. This uncertainty-driven sampling maximizes the information gain per labeled example, efficiently expanding the model's competence to rare or edge-case modulation types.

Epistemic uncertainty spikes when a model encounters inputs from regions of the feature space sparsely covered by training data. This makes it a powerful signal for out-of-distribution (OOD) detection in signal classification.

• Mechanism: A Bayesian neural network or deep ensemble will exhibit high variance in its predictions for unknown modulation types, while showing low variance for known classes like QPSK or 16QAM.
• Application: A spectrum monitoring system can use this uncertainty spike to flag a novel radar waveform for analyst review instead of forcing a confident but incorrect classification.
• Key Distinction: Unlike aleatoric uncertainty (noise), this uncertainty is reducible—collecting more labeled examples of the novel signal would shrink it.

Epistemic uncertainty directly informs active learning loops, where a classifier intelligently selects which unlabeled signals to send to a human annotator for labeling.

• Query Strategy: The model ranks captured signals by their epistemic uncertainty score. Those with the highest uncertainty—likely representing rare or new modulation schemes—are prioritized for expert review.
• Efficiency Gain: This avoids wasting human effort on common, well-understood signals (e.g., FM broadcasts) and focuses annotation budget on the long tail of the signal environment.
• Continuous Adaptation: As newly labeled examples are added to the training set, the model's epistemic uncertainty for that signal type decreases, and the classifier expands its known repertoire.

Bayesian neural networks (BNNs) place probability distributions over model weights instead of learning fixed point estimates, enabling direct quantification of epistemic uncertainty.

• Monte Carlo Dropout: A practical approximation that performs multiple stochastic forward passes at inference time. The variance across these passes for an IQ sample indicates epistemic uncertainty.
• Deep Ensembles: Training multiple independent models with different initializations. For a known signal like BPSK, all models agree; for an unknown OFDM variant, their predicted class probabilities diverge sharply.
• Output: A calibrated uncertainty heatmap over the constellation diagram, showing precisely where in the feature space the model lacks knowledge.

The OpenMax layer replaces a standard SoftMax classifier by explicitly modeling epistemic uncertainty using Extreme Value Theory (EVT).

• Process: For each known class, a Weibull distribution is fitted to the distance between correct training samples and their class mean activation vector. This models the tail behavior of what is 'known'.
• Recalibration: At test time, the activation vector for a new signal is recalibrated. If the signal falls deep in the tail of every known class's Weibull distribution, its probability mass is reassigned to a new 'unknown' pseudo-class.
• Result: The model outputs an explicit probability that the signal is from an unknown modulation scheme, directly driven by a statistical model of its own knowledge gaps.

Epistemic uncertainty can be visualized in the model's embedding space to diagnose blind spots and guide data collection strategies.

• Technique: Using dimensionality reduction (t-SNE or UMAP) on the penultimate layer embeddings, points are colored by their epistemic uncertainty score.
• Interpretation: Tight clusters of known modulations (e.g., 64QAM) appear as low-uncertainty islands. Sparse, isolated points between clusters represent high epistemic uncertainty—potential unknown signal types.
• Engineering Value: This reveals exactly which regions of the signal parameter space (e.g., specific baud rates, carrier frequencies) are underrepresented in the training corpus, directing targeted data acquisition.

Evidential neural networks predict the parameters of a Dirichlet distribution over class probabilities, treating classification as subjective opinion rather than a point estimate.

• Uncertainty Decomposition: The framework naturally decomposes total predictive uncertainty into aleatoric (data noise) and epistemic (model knowledge gaps) components.
• Signal Classification Use: When processing a burst transmission, the model outputs not just 'QPSK with 85% probability' but also a measure of the evidence supporting that claim. Low evidence for all known classes signals epistemic uncertainty.
• Advantage: This provides a principled, single-forward-pass method for detecting unknown modulations without needing multiple models or Monte Carlo sampling.