Open Set Recognition

A statistical framework for modeling the tail of the activation score distribution of known classes. Instead of applying an arbitrary threshold, EVT fits a Generalized Pareto Distribution or Weibull distribution to the extreme values of the distance between a sample and its nearest class centroid. This provides a mathematically principled, per-class probability that a sample is an outlier, enabling calibrated open-set rejection.

A direct replacement for the standard SoftMax layer in a neural network. OpenMax recalibrates the final activation vector by fitting a Weibull model to the distance of each correctly classified training sample from its class mean. During inference, it uses these models to estimate the probability that an input belongs to an unknown class, redistributing the SoftMax probability mass to an explicit 'unknown' pseudo-class.

A representation learning approach where the model learns a reciprocal point for each known class—a latent representation that is maximally separated from the class's training samples. The probability of a sample belonging to a class is inversely proportional to its distance from the reciprocal point. Unknown samples are naturally mapped to a bounded region near the reciprocal points, creating a clear, separable open space risk.

A data augmentation strategy that trains a GAN to generate realistic, synthetic signal samples that lie just outside the known class decision boundaries. By exposing the classifier to these 'boundary' samples during training and labeling them as 'unknown,' the model learns a tighter, more conservative decision boundary. This directly addresses the problem of open space risk without requiring real unknown emitter data.

A non-parametric approach that uses a Siamese network or prototypical network to map I/Q samples into a high-dimensional embedding space. During enrollment, a centroid is computed for each known device. At inference, the Euclidean distance or cosine similarity to the nearest centroid is calculated. If the distance exceeds a calibrated threshold, the sample is rejected as unknown. This method naturally supports few-shot enrollment.

A hybrid technique that first trains a deep clustering model to group known emitters into compact, well-separated clusters. A secondary isolation forest or Local Outlier Factor (LOF) algorithm is then applied to the learned feature space. This secondary model scores each new sample based on its local density deviation, flagging points that fall in low-density regions between clusters as novel, unknown emitters.