Parallel and adaptive signal processing
AssigneeSparsense, Inc.
InventorsZsolt Palotai, András Lörincz
Life of this patent
Estimating the term…
Abstract
A method and apparatus for parallel and adaptive signal reconstruction from a multitude of signal measurements. Algorithms and hardware are disclosed to denoise the measured signals, to compress the measured signals, and to reconstruct the signal from fewer measurements than standard state-of-the-art methods require. A parallel hardware design is disclosed in which the methods that are described can be efficiently executed.
Claims81
A method for parallel adaptive signal reconstruction from a multitude of measurements of an input signal, the method comprising the steps of:
- (a) establishing sparse transforms Q and W to provide a sparse representation of the input signal, wherein the sparse representation of the input signal includes active components of the sparse representation;
- (b) establishing a sparsity measure based on a number of active components of the sparse representation;
- (c) using the sparsity measure to establish a sparsity constraint determining whether a sparse code is sparse enough by the sparsity measure;
- (d) establishing a cost function containing the sparsity measure and a correlation quality of reconstruction of the input signal;
- (e) receiving an input signal in the form of n dimensional sampled values;
- (f) calculating a first sparse code by a selection based method (SM), wherein the first sparse code is calculated using non-zero sampled values of the input signal;
- (g) sparsification of said first sparse code;
- (h) calculating using one or more hardware processor a usage rate of the active components of said sparse code;
- (i) calculating using the one or more hardware processor similarity and dissimilarity measures between the active components of said sparse code;
- (j) determining a learning rate for the active components based on the usage rate and similarity and dissimilarity measures of the active components of said sparse code;
- (k) updating said transforms Q and W with said learning rate to decrease said cost function;
- (l) tuning of said transforms Q and W so that the low usage rate components of said sparse code become more frequent;
- and
- (m) removing values below a predetermined threshold from said transforms Q and W and normalization of said transform Q.
A method of calculating a sparse code by an iterative selection based method (SM) for selecting active components of the sparse code, where said iterative selection based method (SM) comprises the steps of:
- (a) establishing a starting preference for each component of the sparse code to become active;
- (b) modifying preferences based on similarity and dissimilarity measures of components of the sparse code;
- (c) generating at least one phenotype as a population of phenotypes from said modified preferences;
- (d) calculating using one or more hardware processor a cost function of said phenotypes of said population;
- (e) establishing an elite set of said phenotypes, containing a given percent of the best phenotypes according to the cost function;
- (f) updating the preferences of the components based on said elite set;
- and
- (g) repeating the steps (b)-(f) of said selection based method (SM) until the cost function of the best phenotype is below a given threshold and a sparsity measure of said phenotype satisfies a sparsity constraint or reaches a maximal repetition number.
Dependencies are inferred from the claim text.