OpenMax

The core innovation of OpenMax is fitting a Weibull distribution to the tail of the distance distribution between correctly classified training samples and their class mean activation vectors. This statistical model captures the extreme value behavior of each known class, enabling the system to quantify how likely a new activation score is to belong to that class's tail. The Weibull parameters—scale, shape, and location—are computed per class during a calibration phase after standard training completes.

OpenMax does not simply threshold SoftMax probabilities. Instead, it recalibrates the activation vector before applying SoftMax. For each known class, the algorithm estimates a weight based on the fitted Weibull CDF. The top-k activations are then adjusted downward proportionally to their tail probability. A new unknown class activation is synthesized from the residual, creating a k+1 dimensional vector. After recalibration, SoftMax produces a normalized probability distribution that includes an explicit score for the unknown class.

Unlike simple confidence thresholding, OpenMax provides a principled rejection mechanism without requiring manual tuning of a sensitivity parameter. The probability assigned to the synthesized unknown class directly represents the model's belief that the input is novel. If this unknown probability exceeds the maximum known class probability, the sample is rejected. This approach naturally adapts to the geometry of the learned feature space rather than relying on an arbitrary global threshold.

OpenMax implements the concept of meta-recognition—a system that reasons about its own recognition decisions. The Weibull fitting process models the probability of extreme events in the classifier's internal representation, effectively giving the network a statistical model of its own failure modes. This meta-cognitive layer allows the system to recognize when it is operating outside its domain of competence, a critical capability for autonomous spectrum monitoring systems that encounter unknown modulation schemes.

The original OpenMax uses Mean Activation Vectors computed per class from correctly classified training examples. However, the framework generalizes to other distance metrics. Variants include using Euclidean distance to class centroids in the penultimate layer, cosine similarity for angular separation, or Mahalanobis distance to account for class covariance structure. Each variant changes the geometry of the rejection boundary, with Mahalanobis distance providing the most statistically informed separation for Gaussian-distributed features.

OpenMax operates as a drop-in replacement for the final classification layer of any deep neural network. The underlying feature extractor is trained normally with cross-entropy loss. After training, the penultimate layer activations for all correctly classified training samples are collected. The Weibull fitting and activation recalibration logic are then applied as a post-processing step. This modular design allows OpenMax to enhance existing modulation recognition architectures without requiring retraining or modifying the backbone network.

The algorithmic process that underpins OpenMax. Meta-recognition is a system's ability to analyze its own decision-making to determine if an input is outside its learned competence. In OpenMax, this involves:

• Computing the activation vector from the penultimate layer.
• Measuring the distance to the nearest class mean.
• Using the EVT-fitted Weibull CDF to compute a per-class probability of being an outlier. This self-assessment directly recalibrates the final SoftMax scores, enabling rejection of unknown modulations.

A critical step in the OpenMax algorithm. For each known class, a Weibull distribution is fitted to the tail of the distances between correctly classified examples and their class mean. Key parameters include:

• Scale (η): Governs the spread of the distribution.
• Shape (κ): Determines tail heaviness. These parameters are used to compute a CDF probability for each test sample's distance. A high CDF value indicates the sample is likely an outlier relative to that class, directly contributing to the recalibrated 'unknown' probability score.

The overarching classification paradigm that OpenMax was designed to address. Unlike the closed-set assumption, OSR requires a model to both accurately classify known classes and detect samples from unknown classes not seen during training. OpenMax achieves this by augmenting a standard neural network with a meta-recognition step that estimates open space risk. This makes it a foundational algorithm for cognitive radio systems that must operate in dynamic, unpredictable electromagnetic environments.

The core computational mechanism of OpenMax. Instead of directly applying SoftMax to the penultimate layer's logits, OpenMax performs a three-step recalibration:

1. Rank the top-k activation scores.
2. Weight each score by 1 - ω(x), where ω(x) is the Weibull CDF probability that the sample is an outlier for that class.
3. Create a new 'unknown' logit from the sum of the redistributed mass. This process explicitly models the probability of an input belonging to a background or unknown class, which standard SoftMax ignores.

A closely related task often confused with OSR. OOD detection focuses on identifying inputs that differ from the training distribution as a whole, without necessarily distinguishing between known classes. OpenMax bridges OOD and OSR by performing class-conditional outlier detection. It doesn't just flag a signal as anomalous; it estimates the probability that it belongs to a specific unknown class, making it more granular than standard OOD methods like ODIN or Mahalanobis distance-based detectors.