Abstract: Deep RCA uses a modified recursive least squares (RLS) optimization method and a novel null-class vector that together allow the algorithm to remember prior classes as it learns the new class. Deep RCA only has to be trained on the new class data which results in a significant improvement in training speed and almost no memory requirements to achieve the goal of near, real-time class augmentation for deep neural networks.

Abstract: Complex image data is transformed to a holographic representation of the image. A subset of the holographic representation is modeled. Model parameters constitute a compressed image representation. A two-dimensional Fourier transform can be applied to obtain the holographic image. Modeling includes applying an analysis portion of an adaptive analysis/synthesis linear prediction methodology to a subset of the holographic representation to obtain an autoregressive model. Prior to modeling, one-dimensional Fourier transform can be performed on the holographic representation and the linear prediction is one-dimensional. Model parameters are preferably quantized. Embodiments include determining error between the model and the model's input data. There the compressed image representation the error, which also can be quantized. The subset of the holographic representation can be less than all the representation. The subset can be a plurality of complete rows; preferably substantially symmetric about 0 Hz.

Abstract: Embodiments of the invention view image intensity data as a spectrum of underlying wave forms. The spectrum of these waves can be approximated using spectral estimation techniques where the spectrum parameters constitute the compressed image. The image's underlying wave forms can be recovered using an inverse Fourier transform. The original image can also be symmetrically extended prior to the transform to preserve real valued transformed data and model parameters. The modeling method is typically based on a linear predictive methodology to obtain the spectrum parameters. Other transforms include a 2-D Fourier transform that transforms the image into a holographic representation similar to synthetic aperture radar (SAR) phase history. This 2-D waveform holographic format can be decorrelated into 1-D planar waves by applying a 1D Fourier transform. This process enables 1-D linear predictive modeling to obtain the spectral parameters. For compression applications the model parameters are preferably quantized.

Abstract: Image pixel intensity data is transformed to a holographic representation of the image. A subset of the holographic representation is modeled. Model parameters constitute a compressed image representation. A two-dimensional Fourier transform can be applied to obtain the holographic image. Modeling includes applying an analysis portion of an adaptive analysis/synthesis prediction methodology to a subset of the holographic representation. Linear prediction can be the adaptive analysis/synthesis prediction methodology. Prior to modeling, one-dimensional Fourier transform can be performed on the holographic representation and the linear prediction is one-dimensional. Model parameters are preferably quantized. Embodiments include determining error between the model and the model's input data. There the compressed image representation the error, which also can be quantized. The subset of the holographic representation can be less than all the representation.

Abstract: Image pixel intensity data is transformed to a holographic representation of the image. A subset of the holographic representation is modeled. Model parameters constitute a compressed image representation. A two-dimensional Fourier transform can be applied to obtain the holographic image. Modeling includes applying an analysis portion of an adaptive analysis/synthesis prediction methodology to a subset of the holographic representation. Linear prediction can be the adaptive analysis/synthesis prediction methodology. Prior to modeling, one-dimensional Fourier transform can be performed on the holographic representation and the linear prediction is one-dimensional. Model parameters are preferably quantized. Embodiments include determining error between the model and the model's input data. There the compressed image representation the error, which also can be quantized. The subset of the holographic representation can be less than all the representation.

Abstract: Image pixel intensity data is transformed to a holographic representation of the image. A subset of the holographic representation is modeled. Model parameters constitute a compressed image representation. A two-dimensional Fourier transform can be applied to obtain the holographic image. Modeling includes applying an analysis portion of an adaptive analysis/synthesis prediction methodology to a subset of the holographic representation. Linear prediction can be the adaptive analysis/synthesis prediction methodology. Prior to modeling, one-dimensional Fourier transform can be performed on the holographic representation and the linear prediction is one-dimensional. Model parameters are preferably quantized. Embodiments include determining error between the model and the model's input data. There the compressed image representation the error, which also can be quantized. The subset of the holographic representation can be less than all the representation.

Abstract: In a digital signal processor (DSP), input data is configured as a data matrix comprising data samples collected from an input signal. A weight vector is applied to the matrix, where the weight vector comprises three parts including (a) a rank reduction transformation produced by decomposition of data samples in a multistage Wiener filter having a plurality of stages, each stage comprising projection onto two subspaces. Each subsequent stage comprises projecting data transformed by the preceding second subspace onto each of a first subspace comprising a normalized cross-correlation vector at the previous stage and a second subspace comprising the null space of the normalized cross-correlation vector of the current stage, to form a reduced rank data matrix. Part (b) of the weight vector comprises minimizing mean squared error in the reduced rank data space. The output is a linear estimate of input data.

Abstract: In a digital signal processor (DSP), input data is configured as a data matrix comprising data samples collected from an input signal. A weight vector is applied to the data matrix, where the weight vector comprises three parts including (a) a rank reduction transformation produced by decomposition of the data samples in a multistage Wiener filter having a plurality of stages, each stage comprising projection onto two subspaces. Each subsequent stage comprises projecting data transformed by the preceding second subspace onto each of a first subspace comprising a normalized cross-correlation vector at the previous stage and a second subspace comprising the null space of the normalized cross-correlation vector of the current stage, to form a reduced rank data matrix. Part (b) of the weight vector comprises minimizing the mean squared error in the reduced rank data space. The output is a linear estimate of the input data.

Abstract: In a digital signal processor (DSP), input data is configured as a data matrix comprising data samples collected from an input signal. A weight vector is applied to the data matrix, where the weight vector comprises three parts including (a) a rank reduction transformation produced by decomposition of the data samples in a multistage Wiener filter having a plurality of stages, each stage comprising projection onto two subspaces. Each subsequent stage comprises projecting data transformed by the preceding second subspace onto each of a first subspace comprising a normalized cross-correlation vector at the previous stage and a second subspace comprising the null space of the normalized cross-correlation vector of the current stage, to form a reduced rank data matrix. Part (b) of the weight vector comprises minimizing the mean squared error in the reduced rank data space. The output is a linear estimate of the input data.