This technology is a trainable algorithm for optimization-based template search that can be used to efficiently detect and recover signals from noisy high dimensional data.
Signal detection is a crucial task across every field that requires processing sensor data, but the efficient deciphering of these signals remains a challenge. Current methods to detect signals from noisy measurements are based on template matching (or matched filtering). Because these methods find the best matching template by exhaustively correlating each template from a template bank to each input signal, they are computationally inefficient. As the dimensions of input data increase, the template banks required to densely cover signal space can become enormous, causing the detection of signals from high dimensional measurements to become intractable and impractical. As such, there is a need for a scalable signal processing algorithm to detect and recover signal from high dimensional data.
This trainable algorithm detects signal from noisy data by using an unrolled optimization framework that takes advantage of true signals’ low-dimensional structure. By representing templates as a low-dimensional search space and framing the search for the best matching template as an optimization problem, a gradient descent-based solver can rapidly find the best matching template without covering the whole search space, allowing the algorithm to be scalable. To further maximize the efficiency and accuracy of the solver, the algorithm uses unrolled optimization to train the solver’s hyperparameters. Once trained, the model can detect signals efficiently and accurately at deployment.
This technology has been validated with simulated gravitational waveforms and images of randomly transformed handwritten digits 3.
Patent Pending
IR CU24045
Licensing Contact: Greg Maskel