Out-of-Distribution Detection

Identifies when test samples are drawn from a fundamentally different probability distribution than the training data. This is distinct from covariate shift (changes in input distribution) and concept drift (changes in label relationships).

• Detects semantic novelty—inputs belonging to classes unseen during training
• Critical for rejecting unknown modulation schemes in open-set signal recognition
• Differs from anomaly detection by focusing on distribution-level rather than instance-level deviations

A baseline OOD detection method that uses the maximum softmax probability as a confidence score. In-distribution samples typically produce higher maximum probabilities than OOD inputs.

• Simple to implement but often overconfident on OOD samples
• Modern neural networks can produce high-confidence predictions even for unrecognizable inputs
• Improved variants use temperature scaling and energy-based scores for better separation

Uses the Helmholtz free energy of a model's logits as an OOD score, providing a theoretically grounded alternative to softmax confidence. Lower energy indicates in-distribution samples.

• Energy score: E(x) = -T · log(Σ exp(f_i(x)/T)) where T is temperature
• Outperforms softmax thresholding on standard benchmarks
• Aligns with energy-based models and Gibbs distributions in statistical physics

Computes the distance between a test sample's feature representation and the nearest class-conditional Gaussian distribution fitted to training data. OOD samples fall far from all fitted distributions.

• Operates in penultimate layer features rather than logit space
• Requires fitting class-conditional Gaussians with tied or per-class covariance
• Effective with generative classifiers and provides calibrated uncertainty estimates

Leverages the magnitude of gradients computed with respect to a uniform label distribution as an OOD indicator. In-distribution samples produce smaller gradient norms than OOD inputs.

• Uses the KL divergence between model output and uniform distribution
• Gradient norms are larger when the model must significantly adjust parameters to accommodate unfamiliar inputs
• Computationally more expensive than score-based methods

Extends the standard softmax layer with an explicit background class for unknown inputs. Uses Weibull distributions fitted to extreme value theory to model the probability of an input belonging to no known class.

• Recalibrates activation vectors using distance to nearest class mean
• Introduces an explicit unknown class probability in the output layer
• Pioneered open-set recognition in computer vision, now adapted for RF modulation classification