Ati Research, Inc.

Cadence Design Systems, Inc.

Iberium Communications Inc.

Raúl Alejandro Casas

Thomas J Endres

Christopher H Strolle

Samir N Hulyalkar

Samer Lutfi Hijazi

Christopher Rowen

Xuehong Mao

Stephen L Biracree

Rishi Kumar

Piyush Kaul

Azzedine Touzni

Anand M Shah

Slobodan Simovich

Stephen Biracree

Azzédine Touzni

Haosong Fu

Himanshu Charaya

Some embodiments perform, in a multi-layer neural network in a computing device, a convolution operation on input feature maps with multiple convolutional filters. The convolutional filters have multiple filter precisions. In other embodiments, electronic design automation (EDA) systems, methods, and computer-readable media are presented for adding such a multi-layer neural network into an integrated circuit (IC) design.

Filtering in trainable networks

Some embodiments perform, in a multi-layer neural network in a computing device, optimization of the multi-layer neural network, for example by making a convolutional change with a first plurality of convolutional filters, or by making a connection change of a first plurality of convolutional filters. In other embodiments, electronic design automation (EDA) systems, methods, and computer-readable media are presented for adding such a multi-layer neural network into an integrated circuit (IC) design.

Complexity optimization of trainable networks

Aspects of the present disclosure involve a transform domain regression convolutional neural network for image segmentation. Example embodiments include a system comprising a machine-readable storage medium storing instructions and computer-implemented methods for classifying one or more pixels in an image. The method may include analyzing the image to estimate one or more transform domain coefficients using a multi-layered function such as a convolutional neural network. The method may further include generating a segmented image by applying a change of basis transformation to the estimated one or more transform domain coefficients.

Transform domain regression convolutional neural network for image segmentation

The current application is directed to maintaining the correct number of symbols in a protocol frame in a digital communications receiver, to prevent catastrophic failure due to dynamic multipath or cycle slips. Timing recovery and framing are coherent, facilitated by placing channel estimation directly into a larger timing recovery loop.

Coherent synchronization and framing in a digital television receiver

Certain embodiments of the present invention encapsulate a variety of data, for example, FIC data, together with ATSC-M/H Service data, into a single IP stream, to simplify the interface between demodulator and media processor. A single data interface protocol within the receiver can then be used, facilitating integration.

Simplified data interface protocol for digital television receiver

A single-axis receiver processes, for example, a complex vestigial sideband (VSB) modulated signal with an equalizer accounting for DC offset within the modulated signal. The DC offset embedded in a received signal may degrade the process of equalization. The equalizer includes a method of blind estimation of the DC offset for removing the DC offset. The estimate is generated by jointly minimizing a Constant Modulus (CM) cost function over equalizer parameters and a DC offset estimate. An arbitrary DC offset the CM cost function admits local spurious minima in terms of the equalizer function. If the DC offset at the receiver is equal to the DC offset inserted at the transmitter, which is equivalent to neglecting the ISI caused by the channel, then only half of the CM equalizer minima remain unchanged. Provided a power constraint on the signal space is satisfied, the DC offset estimate, computed with a joint optimization of the CM cost function, converges to the desired solution, and, thus, the equalizer parameters are guaranteed to converge to the desired solution.

DC offset correction for constant modulus equalization

Carrier phase recovery employs a single-axis blind cost criterion from the Bussgang class of functions, and its stochastic gradient, to generate a carrier phase error used to adjust a received and demodulated signal to near baseband. For one implementation, the estimate is derived in accordance with a Single-Axis Constant Modulus (SA-CM) criterion and its stochastic gradient via a SA-CM algorithm (SA-CMA). The carrier phase error is then used to adjust the carrier frequency and phase of the received and demodulated signal toward the frequency and phase of the carrier used to modulate the transmitted symbols, driving the carrier phase error to zero. The values used for the phase recovery may be either i) an IIR filtered signal, ii) a processed signal (e.g., decisions for the signal symbols), or iii) an equalized and processed signal.

Carrier phase estimation based on single-axis constant modulus cost criterion and Bussgang criteria

A single-axis receiver processing, for example, complex vestigial sideband modulated signals with an equalizer with forward and feedback filters. Forward and feedback filters have parameters that are initialized and adapted to steady state operation. Adaptive equalization employs linear predictive filtering and error term generation based on various cost criteria. Adaptive equalization includes recursive update of parameters for forward and feedback filtering as operation changes between linear and decision-feedback equalization of either single or multi-channel signals. An adaptive, linear predictive filter generates real-valued parameters that are employed to set the parameters of the feedback filter. In an initialization mode, filter parameters are set via a linear prediction filter to approximate the inverse of the channel's impulse/frequency response and a constant modulus error term for adaptation of the filter parameters. In an acquisition mode, equalization is as linear equalization with a constant modulus error term, and possibly other error terms in combination, for adaptation of the filter parameters. In a tracking mode, equalization is as decision feedback equalization with decision-directed error terms for adaptation of the filter parameters. For some equalizer configurations, feedback filtering is applied to real-valued decisions corresponding to complex-valued received data, and includes real-part extraction of the error term employed for recursive update of filtering parameters. Where a training sequence is available to the receiver, initial parameters for forward filtering are estimated by correlation of the received signal with the training sequence.

Linear prediction based initialization of a single-axis blind equalizer for VSB signals

Symbol timing recovery employs a blind cost criterion from the Bussgang class of functions, and its stochastic gradient, to generate a timing phase error used to adjust sampling of received symbols. For one implementation, the estimate is derived in accordance with the Constant Modulus (CM) criterion and its gradient via the CM algorithm (CMA), and the estimate is calculated from a sequence of samples. This estimate is then used to adjust the period and phase of the sample sequence toward the period and phase of the transmitted symbols, driving the timing phase error to zero. The values used may be either i) samples themselves, ii) processed (e.g., interpolated) samples, or iii) equalized and processed samples. In addition, timing phase error estimates for other cost criteria, including the least mean squares algorithm, may be generated. These timing phase error estimates are selected either alone or in combination for deriving the timing phase error used to adjust the period and phase of the sample sequence.

Growing community of inventors

Doylestown, PA, United States of America

Raúl Alejandro Casas