Evidence Deep Learning

The core innovation is replacing the final SoftMax layer with an activation function (e.g., softplus or exponential) that outputs the concentration parameters (α) of a Dirichlet distribution. Instead of a single probability vector, the network predicts a distribution over distributions. The total evidence, calculated as the sum of concentration parameters, directly quantifies the model's confidence. A high evidence total indicates strong support for a known class, while a low total signals high epistemic uncertainty—a key indicator of an unknown modulation scheme.

This framework formalizes predictions as subjective opinions using Dempster-Shafer theory. The model's output is decomposed into:

• Belief Mass: The probability assigned to each known class, derived from the Dirichlet evidence.
• Uncertainty Mass: A scalar value representing the vacuity of the prediction, calculated as the total number of classes divided by the total evidence. This uncertainty mass is the direct signal for open set recognition, allowing the system to explicitly state 'I don't know' when encountering a novel modulation type, rather than forcing a spurious classification.

Evidence Deep Learning uniquely separates two critical types of uncertainty:

• Epistemic Uncertainty: The model's ignorance due to lack of knowledge, reflected in low total evidence. This is reducible with more training data and is the primary metric for detecting out-of-distribution signals.
• Aleatoric Uncertainty: The inherent, irreducible noise in the data (e.g., low SNR), captured by the spread of the Dirichlet distribution. By monitoring both, a cognitive radio system can distinguish between a genuinely new modulation (high epistemic, low aleatoric) and a known modulation in a noisy channel (low epistemic, high aleatoric), enabling more intelligent spectrum management.

Training a Dirichlet network requires a specialized loss function that fits the concentration parameters to the data. The standard approach minimizes the negative log-likelihood of the Dirichlet distribution, known as a Type II Maximum Likelihood loss. This is often regularized with a Kullback-Leibler (KL) divergence term that penalizes misleading evidence for incorrect classes. The KL term forces the network to produce zero evidence for non-target classes, ensuring that high uncertainty is a true signal of novelty rather than a byproduct of conflicting evidence.

To sharpen the distinction between known and unknown classes, training can be augmented with outlier exposure. An auxiliary dataset of diverse background signals is introduced. The loss function is modified to force the network to produce a uniform Dirichlet prior (zero evidence) for these outlier samples. This contrastive approach teaches the model to actively map unknown signals to a high-uncertainty, low-evidence region of the simplex, creating a more robust rejection boundary than standard SoftMax-based Out-of-Distribution Detection methods like ODIN.

At inference time, novelty detection is a simple, interpretable operation. The model's predicted uncertainty mass is compared against a calibrated threshold. If the uncertainty exceeds the threshold, the sample is rejected as unknown. This threshold can be tuned to balance the Open Set Classification Rate, trading off between known-class accuracy and unknown-class recall without retraining the model. This direct access to a rejection score is a significant advantage over methods that require post-hoc calibration of SoftMax probabilities.

The mathematical backbone of Evidence Deep Learning. Instead of outputting a single point estimate of class probabilities, the model places a Dirichlet distribution over the probability simplex. This distribution's concentration parameters represent the evidence collected for each class. A high concentration indicates strong, decisive evidence, while a uniform Dirichlet with low total evidence represents high epistemic uncertainty—the model's way of saying it lacks sufficient data to decide. The Dirichlet is the conjugate prior of the categorical distribution, enabling closed-form Bayesian updates.

A formal reasoning framework that treats neural network predictions as subjective opinions rather than objective probabilities. An opinion is a triplet consisting of belief mass, uncertainty mass, and a base rate. Evidence Deep Learning directly maps to subjective logic: the Dirichlet parameters define the belief masses for each class, and the remaining probability density represents vacuity—uncertainty due to lack of evidence. This formalism allows for principled opinion fusion when combining predictions from multiple sensors or receivers in a cognitive radio network.

Evidence Deep Learning explicitly decomposes predictive uncertainty into two distinct types:

• Epistemic Uncertainty: The model's ignorance due to insufficient knowledge or training data. This is reducible—collecting more labeled examples of a rare modulation scheme will decrease it. High epistemic uncertainty is the primary signal for novelty detection.
• Aleatoric Uncertainty: The inherent, irreducible noise in the data itself, such as thermal noise or fading in a wireless channel. Evidence-based models capture this through the spread of the Dirichlet distribution, distinguishing a genuinely ambiguous signal from one that is simply unfamiliar.

Evidence Deep Learning provides multiple scalar metrics derived from the Dirichlet output to drive decision-making:

• Total Evidence (S): The sum of all concentration parameters. A low S indicates high overall uncertainty.
• Vacancy/Vacuity: The uncertainty mass in the subjective logic opinion, computed as K/S where K is the number of classes. This directly quantifies the lack of evidence.
• Dissonance: Measures the conflict between belief masses when evidence is distributed across multiple classes, indicating a genuinely ambiguous signal rather than an unknown one. These metrics enable a threshold-based rejection policy for open set recognition.

Training an evidence-based model requires specialized loss functions that learn to output Dirichlet parameters:

• Type II Maximum Likelihood: Minimizes the negative log of the marginal likelihood integrated over the Dirichlet prior, fitting the distribution to the data.
• Bayes Risk with Cross-Entropy: Minimizes the expected categorical cross-entropy under the Dirichlet, providing a more stable training signal.
• KL Divergence Regularization: A critical term that penalizes evidence for incorrect classes, forcing the model to produce a uniform Dirichlet (high uncertainty) for out-of-distribution inputs. Without this, the model can collapse to generating infinite evidence for the wrong class.

While both methods quantify uncertainty, they differ fundamentally:

• Bayesian Neural Networks (BNNs): Place distributions over the model's weights, requiring multiple stochastic forward passes (Monte Carlo dropout or variational inference) to estimate predictive variance. This is computationally expensive at inference time.
• Evidence Deep Learning: Places a distribution directly over the output probabilities in a single deterministic forward pass. The uncertainty is encoded in the Dirichlet parameters, making it far more efficient for real-time spectrum classification on edge hardware. It is a form of prior networks that learns the distribution of predictions rather than the distribution of weights.