Abstract: A reorganizable neural network computing device is provided. The computing device includes a data processing array unit including a plurality of operators disposed at locations corresponding to a row and a column. One or more chaining paths which transfer the first input data from the operator of the first row of the data processing array to the operator of the second row are optionally formed. The plurality of first data input processors of the computing device transfer the first input data for a layer of the neural network to the operators along rows of the data processing array unit, and the plurality of second data input processors of the computing device transfer the second input data to the operators along the columns of the data processing array.

Abstract: A point estimate value for an individual is computed using a Bayesian neural network model (BNN) by training a first BNN model that computes a weight mean value, a weight standard deviation value, a bias mean value, and a bias standard deviation value for each neuron of a plurality of neurons using observations. A plurality of BNN models is instantiated using the first BNN model. Instantiating each BNN model of the plurality of BNN models includes computing, for each neuron, a weight value using the weight mean value, the weight standard deviation value, and a weight random draw and a bias value using the bias mean value, the bias standard deviation value, and a bias random draw. Each instantiated BNN model is executed with the observations to compute a statistical parameter value for each observation vector of the observations. The point estimate value is computed from the statistical parameter value.

Abstract: Embodiments described herein combine both masked reconstruction and predictive coding. Specifically, unlike contrastive learning, the mutual information between past states and future states are directly estimated. The context information can also be directly captured via shifted masked reconstruction—unlike standard masked reconstruction, the target reconstructed observations are shifted slightly towards the future to incorporate more predictability. The estimated mutual information and shifted masked reconstruction loss can then be combined as the loss function to update the neural model.

Abstract: A design optimization method based on active learning, which involves dynamic exploration and exploitation of the design space of interest using an ensemble of machine learning algorithms. In this approach, a hybrid methodology incorporating an explorative weak learner which fits high-level information about the response surface, and an exploitative strong learner (based on committee machine) that fits finer details around promising regions identified by the weak learner, is employed. For each design iteration, an aristocratic approach is used to select a set of nominees, where points that meet a threshold merit value as predicted by the weak learner are selected to be evaluated using function evaluation. In addition to these points, the global optimum as predicted by the strong learner is also evaluated to enable rapid convergence to the actual global optimum once the most promising region has been identified by the optimizer.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training an action selection neural network. One of the methods includes maintaining a replay memory that stores trajectories generated as a result of interaction of an agent with an environment; and training an action selection neural network having policy parameters on the trajectories in the replay memory, wherein training the action selection neural network comprises: sampling a trajectory from the replay memory; and adjusting current values of the policy parameters by training the action selection neural network on the trajectory using an off-policy actor critic reinforcement learning technique.

Abstract: Networks and encodings therefor are provided that are adapted to provide increased energy efficiency and speed for convolutional operations. In various embodiments, a neural network comprises a plurality of neural cores. Each of the plurality of neural cores comprises a memory. A network interconnects the plurality of neural cores. The memory of each of the plurality of neural cores comprises at least a portion of a weight tensor. The weight tensor comprising a plurality of weights. Each neural core is adapted to retrieve locally or receive a portion of an input image, apply the portion of the weight tensor thereto, and store locally or send a result therefrom via the network to other of the plurality of neural cores.

Abstract: Disclosed is a method for convolution calculation in a neural network, comprising: reading an input feature map, depthwise convolution kernels and pointwise convolution kernels from a dynamic random access memory (DRAM); performing depthwise convolution calculations and pointwise convolution calculations by depthwise convolution calculation units and pointwise convolution calculation units, according to the input feature map, the depthwise convolution kernels and the pointwise convolution kernels to obtain output feature values of a first predetermined number p of points on all pointwise convolution output channels; storing the output feature values of a first predetermined number p of points on all pointwise convolution output channels into an on-chip memory; and repeating above operation to obtain output feature values of all points on all point wise convolution output channels. Therefore, the storage space for storing intermediate results may be reduced.

Abstract: Some embodiments provide a compiler for optimizing the implementation of a machine-trained network (e.g., a neural network) on an integrated circuit (IC). The compiler of some embodiments receives a specification of a machine-trained network including multiple layers of computation nodes and generates a graph representing options for implementing the machine-trained network in the IC. In some embodiments, the graph includes nodes representing options for implementing each layer of the machine-trained network and edges between nodes for different layers representing different implementations that are compatible. In some embodiments, the graph is populated according to rules relating to memory use and the numbers of cores necessary to implement a particular layer of the machine trained network such that nodes for a particular layer, in some embodiments, represent fewer than all the possible groupings of sets of clusters.

Abstract: In some examples, an individually-pruned neural network can estimate blood pressure from a seismocardiogram (SCG). In some examples, a baseline model can be constructed by training the model with SCG data and blood pressure measurement from a plurality of subjects. One or more filters (e.g., the filters in the top layer of the network) can be ranked by separability, which can be used to prune the model for each unseen user that uses the model thereafter, for example. In some examples, individuals can use individually-pruned models to calculate blood pressure using SCG data without corresponding blood pressure measurements.

Abstract: A technology for obtaining a respiratory rate from a photoplethysmogram (PPG) signal. In one example, an artificial neural network model can be trained to predict a respiratory rate using a training dataset containing PPG data. The artificial neural network model can include a first series of convolutional layers to remove artifacts from a PPG signal, a fast Fourier transform (FFT) layer to convert the PPG signal to PPG frequency representations, and a dense layer to decode the PPG frequency representations to respiratory rate predictions. After training the artificial neural network model, PPG data generated by a pulse oximeter monitor can be obtained, and the PPG data can be input to the artificial neural network model. The artificial neural network model outputs a respiratory rate prediction, wherein the respiratory rate prediction represents the respiratory rate obtained from the PPG signal.

Abstract: Methods and apparatus are provided for real-time anomaly detection over sets of time-series data. One method comprises: obtaining a state-space representation of a plurality of states and transitions between said states based on sets of historical time-series data; obtaining an anomaly detection model trained using a supervised learning technique, wherein the anomaly detection model associates sequences of states in the state-space representation with annotated anomalies in the sets of historical time-series data and assigns a probability to said sequences of states; and, for incoming real-time time-series data, determining a likelihood of a current state belonging to a plurality of possible states in the state-space representation; and determining a probability of incurring said annotated anomalies based on a plurality of likely current state sequences that satisfy a predefined likelihood criteria.

Abstract: A system for generating an alimentary instruction set identifying an individual prognostic mitigation plan, the system comprising a computing device; a diagnostic engine operating on the computing device, wherein the diagnostic engine is configured to receive information related to a biological extraction of a user, wherein the biological extraction contains an element of user physiological state data; generate a diagnostic output based upon the information related to the biological extraction, and an alimentary instruction set generator module operating on the computing device, wherein the alimentary instruction set is configured to identify an element of user wellness behavior data; generate a nutrition instruction set utilizing the diagnostic output, the element of user wellness behavior data and a first machine-learning process; and customize the ameliorative output to identify a prognostic mitigation plan utilizing the element of user wellness data, and the nutrition instruction set.

Abstract: Systems and methods are provided for optimizing GPU memory allocation for high-performance applications such as deep learning (DL) computing. For example, a DL task is executed using GPU resources (GPU device and GPU memory) to process a DL model having functional layers that are processed in a predefined sequence. A current functional layer of the DL model is invoked and processed using the GPU device. In response to the invoking, a data compression operation is performed to compress data of a previous functional layer of the DL model, and store the compressed data in the GPU memory. Responsive to the invoking, compressed data of a next functional layer of the DL model is accessed from the GPU memory and a data decompression operation is performed to decompress the compressed data for subsequent processing of the next functional layer of the DL model by the GPU device.

Abstract: A device, system, and method for approximating a neural network comprising N synapses or filters. The neural network may be partially-activated by iteratively executing a plurality of M partial pathways of the neural network to generate M partial outputs, wherein the M partial pathways respectively comprise M different continuous sequences of synapses or filters linking an input layer to an output layer. The M partial pathways may cumulatively span only a subset of the N synapses or filters such that a significant number of the remaining the N synapses or filters are not computed. The M partial outputs of the M partial pathways may be aggregated to generate an aggregated output approximating an output generated by fully-activating the neural network by executing a single instance of all N synapses or filters of the neural network. Training or prediction of the neural network may be performed based on the aggregated output.

Abstract: Systems and methods for mitigating defects in a crossbar-based computing environment are disclosed. In some implementations, an apparatus comprises: a plurality of row wires; a plurality of column wires connecting between the plurality of row wires; a plurality of non-linear devices formed in each of a plurality of column wires configured to receive an input signal, wherein at least one of the non-linear device has a characteristic of activation function and at least one of the non-linear device has a characteristic of neuronal function.

Abstract: In one aspect, a method of a neuron circuit includes the step of providing a plurality of 2N?1 single-level-cell (SLC) flash cells for each synapse (Yi) connected to a bit line forming a neuron. The method includes the step of providing an input vector (Xi) for each synapse Yi wherein each input vector is translated into an equivalent electrical signal ESi (current IDACi, pulse, TPULSEi, etc). The method includes the step of providing an input current to each synapse sub-circuit varying from 20*ESi to (2N?1)*ESi. The method includes the step of providing a set of weight vectors or synapse (Yi), wherein each weight vector is translated into an equivalent threshold voltage level or resistance level to be stored in one of many non-volatile memory cells assigned to each synapse (Yi).

Abstract: Systems for predicting communication settlement times across disparate networks store a first tier of a machine learning architecture comprising multiple machine learning models and an aggregation layer; store a second tier comprising rule sets for predicting settlement times; receive multiple data feeds corresponding to multiple communication data types; generate feature inputs based on the data feeds; input the feature inputs into the respective models to generate respective outputs; generate, using the aggregation layer, a third feature input based on the outputs; determine, based on the third feature input, a first rule set for predicting settlement times; receive a communication; predict a settlement time based on the first rule set; determine an aggregated communication load at a first time based on the settlement time; determine a performance availability requirement based on the load; determine a recommendation based on the performance availability requirement; and generate the recommendation based on

Abstract: Systems and methods for projecting one or more trends in electronic data and generating enhanced data. A system includes a data forecasting system is in electronic communication with one or more electronic data sources via an electronic network. The data forecasting system is configured to: monitor the electronic data source(s) for data that meet one or more predetermined criteria; obtain at least a portion of the monitored data from electronic data source(s) based on the predetermined criteria; create a data set from the obtained data; derive one or more data values associated with the data set over a predetermined period according to a forward-looking term methodology; and utilize the data set and the derived value(s) over the predetermined period to derive at least one data forecast metric associated with the data set.

Abstract: The present disclosure relates to a neuron for an artificial neural network, the neuron comprising: a first dot product engine and a second dot product engine. The first dot product engine is operative to: receive a first set of weights; receive a set of inputs; and calculate the dot product of the set of inputs and the first set of weights to generate a first dot product engine output. The second dot product engine is operative to: receive a second set of weights; receive the set of inputs; and calculate the dot product of the set of inputs and the second set of weights to generate a second dot product engine output. The neuron further comprises a combiner operative to combine the first dot product engine output and the second dot product engine output to generate a combined output, and an activation function module arranged to apply an activation function to the combined output to generate a neuron output.

Abstract: A property vector representing extractable measurable properties, such as musical properties, of a file is mapped to semantic properties for the file. This is achieved by using artificial neural networks “ANNs” in which weights and biases are trained to align a distance dissimilarity measure in property space for pairwise comparative files back towards a corresponding semantic distance dissimilarity measure in semantic space for those same files. The result is that, once optimised, the ANNs can process any file, parsed with those properties, to identify other files sharing common traits reflective of emotional-perception, thereby rendering a more liable and true-to-life result of similarity/dissimilarity. This contrasts with simply training a neural network to consider extractable measurable properties that, in isolation, do not provide a reliable contextual relationship into the real-world.