Open Space Risk

Open Space Risk is formally defined as the relative measure of positively labeled open space compared to the overall measure of positively labeled space. Open space is any region in the feature space sufficiently far from known training data. A robust open set classifier must minimize this risk by bounding the positive recognition region tightly around known class clusters, ensuring that the infinite expanse of unknown feature space is not inadvertently labeled as known.

Extreme Value Theory (EVT) provides the statistical foundation for managing Open Space Risk. EVT models the distribution of rare, extreme events—in this context, the maximum distance between a correctly classified sample and its class mean. By fitting a Weibull distribution to these tail distances, the system can compute a calibrated probability of inclusion. This allows the model to reject inputs whose distance from known data is so extreme that they are statistically more likely to be unknown.

Minimizing Open Space Risk creates a fundamental trade-off. A classifier must tightly bound the known class space to reject unknowns, but it must also generalize to legitimate variations of known classes. Overly tight bounds cause false rejections of valid known samples, while loose bounds increase the risk of labeling unknowns as known. Solutions like angular margin losses (e.g., ArcFace) help by maximizing inter-class separation and compacting intra-class variance in the embedding space.

The Openness Measure provides a standardized way to quantify the difficulty of an open set problem and the associated risk. It is defined as:

• Openness = 1 - sqrt( (2 * N_train) / (N_test + N_target) ) Where N_train is the number of known classes used for training, N_test is the number of classes seen at test time, and N_target is the number of classes the system must eventually identify. A higher openness score indicates a larger proportion of unknown classes and, consequently, a greater exposure to Open Space Risk.

Practical rejection of open space relies on distance metrics in the learned feature embedding. Common techniques include:

• Mahalanobis Distance: Accounts for class covariance, creating ellipsoidal bounds.
• Deep SVDD: Trains a neural network to map all known data into a minimal-volume hypersphere, treating anything outside as open space.
• Reconstruction Error: Uses autoencoders; unknown samples yield high reconstruction error as they fall outside the learned manifold, indicating they occupy open space.