Parametric Curve Encoding for Anomaly Detection Model Architecture

The patent applies preliminary action by using parametrically modeled curves (such as quadratic Bézier curves) to pre-define and constrain the search space of possible model architectures. Instead of iteratively testing arbitrary architectures, the method generates candidate architectures that lie on predefined curves in the architecture space, thereby preparing and narrowing down potential solutions before the actual optimization process begins. This reduces the number of iterations needed to find accurate models.

The patent applies parameter changes by representing model architectures through parametric curves where a small number of parameters (control points of the Bézier curve) define the entire architecture structure. By optimizing these curve parameters instead of individually tuning each layer and neuron, the method achieves accurate anomaly detection while maintaining manageable model complexity. The parametric representation allows continuous adjustment of architecture parameters to find the optimal balance between accuracy and complexity.

The patent applies dimensionality change by transforming the discrete, high-dimensional search space of model architectures into a continuous, lower-dimensional curve space. Instead of searching through discrete combinations of layers and neurons, the method represents architectures as points on continuous parametric curves, reducing the search dimensionality. This transformation enables more efficient optimization with fewer iterations, improving productivity while maintaining detection accuracy.