Open Space Risk

Open Space Risk is the theoretical measure of the feature space that is far from any known training data yet is still classified as belonging to a known class. It quantifies the danger of a model extrapolating its decision boundaries into regions of the input space where it has no empirical support. Formally, it is the relative measure of the open space $O$ compared to the total positively labeled space, where $O$ is defined as the space sufficiently distant from all known positive training samples. A robust open-set classifier must minimize this risk by bounding its recognition space tightly around known data.

The danger zone is the specific region of the feature space where open space risk materializes. It is characterized by high-confidence, low-certainty predictions. Key properties include:

• High SoftMax Scores: The model assigns a high probability to a known class despite the input being novel.
• Low Feature Norm: In architectures like Objectosphere, unknown samples often exhibit a smaller feature magnitude.
• Epistemic Uncertainty: The model's ignorance is high, but its aleatoric (data) uncertainty appears low. This combination creates a silent failure where a novel modulation scheme is confidently misidentified as a known one.

Open space risk cannot be measured directly on unknown data but is bounded theoretically. Practical proxies include:

• Open Set Classification Rate (OSCR): Jointly measures accuracy on knowns and correct rejection of unknowns.
• Area Under the ROC Curve (AUROC): Evaluates the binary discrimination between known and unknown samples using a novelty score.
• F1-Score on Rejection: Balances precision and recall for the 'unknown' class. These metrics are computed using a hold-out set of novel modulations not present in the training data, simulating an open-world deployment.

Feature collapse is a primary cause of high open space risk. It occurs when a deep neural network maps all inputs—both known and unknown—to a restricted, dense region of the embedding space. When collapse happens:

• The distance between known class prototypes and unknown samples becomes indistinguishable.
• The model loses its ability to separate novel signals from known ones.
• Open space risk approaches its maximum because the entire unbounded feature space is effectively labeled as known. Preventing collapse through contrastive learning or entropic loss functions is essential for minimizing open space risk.

Extreme Value Theory (EVT) provides a statistical framework to directly model and bound open space risk. The OpenMax algorithm uses EVT by:

1. Computing the distance of every correctly classified training sample to its class mean.
2. Fitting a Weibull distribution to the tail of these distances for each class.
3. Recalibrating the SoftMax probability to include an 'unknown' class based on the cumulative distribution function of the Weibull model. This statistically bounds the recognition space, ensuring that points far from the training distribution are assigned a high probability of being unknown.

The closed-set assumption posits that all test classes are identical to training classes, which implicitly sets open space risk to zero by definition—a dangerous fallacy in dynamic environments. In contrast, open-set recognition acknowledges that:

• The feature space is unbounded and contains infinite unknown classes.
• Labeling any point in this infinite space as 'known' incurs a risk.
• The only way to achieve zero open space risk is to label the entire space as 'unknown' except for a finite, bounded region around known training data. This fundamental shift in perspective is critical for deploying classifiers in real-world spectrum monitoring where new modulation types constantly emerge.

The theoretical risk that an unknown input far from any known training data is nonetheless labeled as a known class. It is quantified as the relative measure of the feature space distant from training samples that is still classified as known. This risk arises because traditional SoftMax classifiers create unbounded decision regions, leaving vast open spaces where the model can make high-confidence but incorrect predictions.

A statistical framework for modeling the tail behavior of distributions, used to bound open space risk. In open set recognition, EVT fits a Weibull distribution to the distance of correct classifications from their class mean. This allows the model to estimate the probability that a query sample belongs to the tail of a known class distribution, providing a principled rejection threshold for unknowns.

A deep learning layer that replaces the standard SoftMax function to explicitly manage open space risk. OpenMax recalibrates activation vectors using Extreme Value Theory to estimate the probability of an input belonging to an unknown class. It fits a Weibull model to the distance between correctly classified training samples and their class mean activation vectors, then uses this to compute a calibrated unknown probability.

A loss function designed to create a distinct separation in feature magnitude between known and unknown samples. It maximizes the feature norm for known classes while minimizing it for unknowns, creating a thresholdable rejection space. This directly addresses open space risk by ensuring that unknown inputs map to a low-magnitude region near the origin, far from the high-magnitude known class representations.

A classification strategy that represents each known class by a reciprocal point in the embedding space rather than a central prototype. The model uses the maximum distance to these reciprocal points to identify unknown samples. By pushing known class embeddings away from their reciprocal points and pulling unknowns toward them, this method creates a bounded, closed decision space that minimizes open space risk.

A failure mode in deep learning where the embeddings of all inputs, including unknowns, map to a restricted region of the feature space. This destroys the model's ability to separate known from novel classes and dramatically increases open space risk. Feature collapse often occurs when a network is trained only on closed-set objectives without explicit regularization to preserve the structure of the embedding space.