Evidential Deep Learning

Instead of outputting a single probability vector, the model predicts the concentration parameters of a Dirichlet distribution. This distribution serves as a higher-order, evidence-based prior over the likelihood of each class.

• The sum of concentration parameters represents total evidence
• High evidence produces a sharp, confident distribution
• Low, uniform evidence indicates high epistemic uncertainty

The framework decomposes the model's output into three interpretable quantities derived from Subjective Logic theory:

• Belief Mass: The probability assigned to each known class based on supporting evidence
• Uncertainty Mass: A reserved probability budget that captures model ignorance
• This explicit uncertainty term spikes for inputs far from the training distribution, enabling native out-of-distribution detection

Training minimizes a specialized loss combining Type II Maximum Likelihood with a regularization term:

• The likelihood term fits the Dirichlet to observed class labels
• A Kullback-Leibler divergence regularizer penalizes misleading evidence for misclassified samples
• This forces the model to shrink evidence to zero for non-target classes, preventing overconfident wrong predictions

The uncertainty mass provides a built-in OOD score without requiring auxiliary outlier datasets or complex post-hoc calibration:

• In-distribution inputs generate high evidence and low uncertainty
• Anomalous or novel inputs produce uniformly low evidence across all classes
• A simple threshold on uncertainty mass or total evidence serves as the rejection criterion

Unlike Monte Carlo Dropout or Deep Ensembles that require multiple stochastic forward passes, evidential deep learning produces uncertainty estimates in a single deterministic forward pass:

• No sampling overhead at inference time
• Computationally efficient for real-time and edge deployment
• Directly models a distribution over distributions rather than sampling from a weight posterior

Evidential networks naturally extend to open set recognition by treating unknown classes as inputs that fail to accumulate evidence for any known category:

• The model can be combined with OpenMax-style rejection layers
• Uncertainty mass replaces the need for Weibull calibration on activation vectors
• Enables robust operation in dynamic electromagnetic environments with previously unseen emitters

The core probability distribution used in evidential deep learning, defined over the simplex of class probabilities. Unlike a single SoftMax point estimate, it parameterizes a second-order probability—a distribution over distributions. The concentration parameters (α) directly encode the evidence collected for each class, allowing the model to express uncertainty through the distribution's sharpness or flatness.

In Subjective Logic, the Dirichlet parameters are decomposed into belief mass (evidence allocated to specific classes) and uncertainty mass (evidence withheld). The total strength S = Σα_k defines the confidence. A high S concentrates belief; a low S (e.g., uniform Dirichlet with α_k = 1) indicates maximum vacuity, signaling an out-of-distribution input to the rejection logic.

The training objective that fits the Dirichlet to the data by maximizing the marginal likelihood of the categorical labels, marginalized over the Dirichlet prior. This contrasts with standard MLE, which fits a point estimate. The resulting loss function, often derived from the Bayesian risk or the negative log of the marginal likelihood, directly incentivizes the network to inflate the evidence for the correct class while shrinking it for others.

A critical regularization term added to the evidential loss to penalize evidence for incorrect classes. The KL divergence between the predicted Dirichlet and a uniform Dirichlet (zero evidence) is minimized for all classes except the ground truth. This forces the model to assign zero evidence to mismatched classes, preventing evidence inflation and ensuring a sharp, well-calibrated belief distribution.

Evidential deep learning provides a natural decomposition of predictive uncertainty. Aleatoric uncertainty (data noise) is captured by the expected categorical entropy under the Dirichlet. Epistemic uncertainty (model ignorance) is captured by the mutual information between the class label and the Dirichlet distribution. This decomposition is crucial for distinguishing ambiguous in-distribution signals from truly unknown emitter types.

The neural network's final layer is modified to output non-negative evidence vectors (e.g., via SoftPlus, ReLU, or exp activation) instead of logits. These evidence values directly parameterize the Dirichlet concentration parameters: α_k = e_k + 1. The additive 1 represents the prior base rate, ensuring the Dirichlet is well-defined even with zero evidence. This simple architectural change transforms a standard classifier into an evidential one.