A method for decentralized multi-task learning includes publishing metadata associated with a first task. A plurality of parameter vectors associated with a set of similar tasks to the first task is obtained and the set of similar tasks is associated with a plurality of other participants. A parameter vector associated with a machine learning dataset for the first task is trained based on a loss function associated with the first task and the plurality of parameter vectors associated with the set of similar tasks. The parameter vector associated with the machine learning dataset for the first task is published.

Described is a cognitive blind source separator (CBSS). The CBSS includes a delay embedding module that receives a mixture signal (the mixture signal being a time-series of data points from one or more mixtures of source signals) and time-lags the signal to generate a delay embedded mixture signal. The delay embedded mixture signal is then linearly mapped into a reservoir to create a high-dimensional state-space representation of the mixture signal. The state-space representations are then linearly mapped to one or more output nodes in an output layer to generate pre-filtered signals. The pre-filtered signals are passed through a bank of adaptable finite impulse response (FIR) filters to generate separate source signals that collectively formed the mixture signal.

A hardware acceleration module may generate a channel-wise argmax map using a predefined set of hardware-implemented operations. In some examples, a hardware acceleration module may receive a set of feature maps for different image channels. The hardware acceleration module may execute a sequence of hardware operations, including a portion(s) of hardware for executing a convolution, rectified linear unit (ReLU) activation, and/or layer concatenation, to determine a maximum channel feature value and/or argument maxima (argmax) value for a set of associated locations within the feature maps. An argmax map may be generated based at least in part on the argument maximum for a set of associated locations.

A real-time multi-channel EEG signal processor based on an on-line recursive independent component analysis is provided. A whitening unit generates covariance matrix by computing covariance according to a received sampling signal. A covariance matrix generates a whitening matrix by a computation of an inverse square root matrix calculation unit. An ORICA calculation unit computes the sampling signal and the whitening matrix to obtain a post-whitening sampling signal. The post-whitening sampling signal and an unmixing matrix implement an independent component analysis computation to obtain an independent component data. An ORICA training unit implements training of the unmixing matrix according to the independent component data to generate a new unmixing matrix. The ORICA calculation unit may use the new unmixing matrix to implement an independent component analysis computation. Hardware complexity and power consumption can be reduced by sharing registers and arithmetic calculation units.

Biometric identification using electroencephalogram (EEG) signals is provided. Embodiments are targeted for biometric applications, where an individual can be identified with a precision of over 99%, using sensed brain signals. In particular, a method is described which can extract unique biomarkers from EEG response signals to classify individuals, also referred to as simple visual reaction task-based EEG biometry (SVRTEB). A subject experiences a simple stimulus or task, and a multi-channel EEG response is recorded. Unique biomarkers are extracted from the recorded EEG response (e.g., as periodogram data points corresponding to different frequencies observed in the brain waves, which can be used to identify a person). A novel signal processing approach uses neural network-based architecture to analyze the EEG response and identify the subject. This signal processing architecture can be readily implemented on hardware and provides high accuracy, precision, and recall.

Methods and systems for combining information present in measured images of semiconductor wafers with additional measurements of particular structures within the measured images are presented herein. In one aspect, an image-based signal response metrology (SRM) model is trained based on measured images and corresponding reference measurements of particular structures within each image. The trained, image-based SRM model is then used to calculate values of one or more parameters of interest directly from measured image data collected from other wafers. In another aspect, a measurement signal synthesis model is trained based on measured images and corresponding measurement signals generated by measurements of particular structures within each image by a non-imaging measurement technique. Images collected from other wafers are transformed into synthetic measurement signals associated with the non-imaging measurement technique and a model-based measurement is employed to estimate values of parameters of interest based on the synthetic signals.

A signal processing device according to an exemplary aspect of the present invention includes: at least one memory storing a set of instructions; and at least one processor configured to execute the set of instructions to: extract, from a target signal, a feature value representing a feature of the target signal; calculate, based on the extracted feature value, signal element bases capable of representing a plurality of object signals by linear combination, and information of the linear combination, weights each representing intensities of the plurality of object signals included in the target signal; derive, based on the weights, information of a target object signal included in the target signal, the target object signal being at least one type of the plurality of object signals; and output information of the target object signal.

The present disclosure describes systems, non-transitory computer-readable media, and methods for accurately and efficiently removing objects from digital images taken from a camera viewfinder stream. For example, the disclosed systems access digital images from a camera viewfinder stream in connection with an undesired moving object depicted in the digital images. The disclosed systems generate a temporal window of the digital images concatenated with binary masks indicating the undesired moving object in each digital image. The disclosed systems further utilizes a generator as part of a 3D to 2D generative adversarial neural network in connection with the temporal window to generate a target digital image with the region associated with the undesired moving object in-painted. In at least one embodiment, the disclosed systems provide the target digital image to a camera viewfinder display to show a user how a future digital photograph will look without the undesired moving object.

The present disclosure describes systems, non-transitory computer-readable media, and methods for accurately and efficiently removing objects from digital images taken from a camera viewfinder stream. For example, the disclosed systems access digital images from a camera viewfinder stream in connection with an undesired moving object depicted in the digital images. The disclosed systems generate a temporal window of the digital images concatenated with binary masks indicating the undesired moving object in each digital image. The disclosed systems further utilizes a generator as part of a 3D to 2D generative adversarial neural network in connection with the temporal window to generate a target digital image with the region associated with the undesired moving object in-painted. In at least one embodiment, the disclosed systems provide the target digital image to a camera viewfinder display to show a user how a future digital photograph will look without the undesired moving object.

A method for classifying input data includes receiving the input data that describe an object, wherein the input data corresponds to plural classes; associating the input data with voxels that describe the object; calculating a real-number sequence X(n), which is associated with a measured parameter P that describes the object; quantizing the real-number sequence X(n) to generate a finite set sequence Q(n), where n describes a number of levels; generating a voxel-based weight matrix for each class of the input data; and calculating a score S for each class of the plural classes, based on a corresponding voxel-based weight matrix. The score S is a number that indicates a likelihood that the input data associated with a given sample belongs to a class of the plural classes.