Abstract: Systems and methods for realigning streams of neuron outputs are provided. An example method may include generating, by a processing unit, neuron outputs including at least a first neuron output and a second neuron output, generating, by at least one further processing unit, further neuron outputs including at least a further first neuron output and a further second neuron output, receiving, by a synchronization module communicatively coupled to the processing unit and the further processing unit, the neuron outputs, wherein the neuron outputs and the further neuron outputs are received in an arbitrary order, and ordering, by the synchronization module, the first neuron output, the further first neuron output, the second neuron output and the further second neuron output according to a further order, the further order being different from the arbitrary order.

Abstract: Systems and methods for performing a quantization of artificial neural networks (ANNs) are provided. An example method may include receiving a description of an ANN and input data associated with the ANN, wherein the input data are represented according to a first data type; selecting a first value interval of the first data type to be mapped to a second value interval of a second data type; performing, based on the input data and the description of the ANN, the computations of one or more neurons of the ANN, wherein the computations are performed for at least one value within the second value interval, the value being a result of mapping a value of the first value interval to a value of the second value interval; determining, a measure of saturations in neurons of the ANN, and adjusting, based on the measure of saturations, the value intervals.

Abstract: Systems and methods for caching data in artificial neural network computations are disclosed. An example method may comprise receiving, by a communication unit, data and a logical address of the data, the data being associated with the ANN, determining, by a processing unit coupled to the communication unit and to a plurality of physical memories and based on the logical address and physical parameters of the physical memories, a physical address of a physical memory of the plurality of physical memories, and performing, by the processing unit, an operation associated with the data and the physical address. The determination of the physical address can be based on a usage count of the data in the ANN computation or a time lapse between a time the data is written to the physical memory and a time the data is used in the ANN computation.

Abstract: Systems and methods for accelerating neuron computations in artificial neural network (ANN) are provided. An example method may comprise receiving, for calculation of a neuron of an ANN, a plurality of first values represented by A bits and a plurality of second values represented by B bits, splitting each value of the plurality of the first values into a set of parts, a count of bits of each of set of parts being less than A, to obtain a set of pluralities of parts, selectively performing mathematical operations on a first plurality of the set of pluralities and the plurality of the second values to obtain a first result, selectively performing further mathematical operations on further pluralities of the set of pluralities and the plurality of the second values to obtain further results, and determining, based on the first result and the further results, an output of the neuron.

Abstract: Systems and methods for accelerating artificial neural network computation are disclosed. An example may comprise selecting, by a controller communicatively coupled to a selector and an arithmetic unit and based on a criterion, an input value from the stream of input values of a neuron, configuring, by the controller, the selector to provide, dynamically, the selected input value to the arithmetic unit, providing, by the controller to the arithmetic unit, an information of the selected input value, acquiring, by the arithmetic unit and based on the information, a weight from a set of weights, and performing, by the arithmetic unit a mathematical operation on the selected input value and the weight to obtain a result, wherein the result is to be used to compute an output of the neuron. The criterion may include a comparison between the input value and a reference value. The reference value may include zero.

Abstract: Systems and methods for performing a quantization of artificial neural networks (ANNs) are provided. An example method may include receiving a description of an ANN and input data associated with the ANN, wherein the input data are represented according to a first data type; selecting a first value interval of the first data type to be mapped to a second value interval of a second data type; performing, based on the input data and the description of the ANN, the computations of one or more neurons of the ANN, wherein the computations are performed for at least one value within the second value interval, the value being a result of mapping a value of the first value interval to a value of the second value interval; determining, a measure of saturations in neurons of the ANN, and adjusting, based on the measure of saturations, the value intervals.

Abstract: Systems and methods for performing a quantization of artificial neural networks (ANNs) are provided. An example method may include receiving a description of an ANN and sets of inputs to neurons of the ANN, the description including sets of weights of the inputs, the weights being of a first data type, determining a first interval of the first data type to be mapped to a second interval of a second data type; performing computations of sums of products of the weights and the inputs to obtain a set of sum results, wherein the computations are performed using at least one number within the second interval, the number being a result of mapping of a number of the first interval to a number of the second interval, determining a measure of saturations in sum results, and adjusting, based on the measure of saturations, one of the first and second intervals.

Abstract: Systems and methods for optimizing artificial neural network (ANN) computations based on automatic determination of a batch size are disclosed. An example method may comprise receiving, by an optimization module, an ANN structure associated with the ANN, and generating, based on the ANN structure, a configuration for a computation engine capable of performing computation of the layers of the ANN. The configuration may include information concerning a batch size of one or more layers of the ANN. The batch size of a layer can be determined based on a bandwidth required to read data related to layer, a number of parameters associated with the layer, and a time the layer processes one input dataset from the batch. The batch size of the layer can differ from the batch size of the ANN. The batch size of the layer may differ from a batch size of another layer of ANN.

Abstract: Systems and methods for caching data in artificial neural network computations are disclosed. An example method may comprise receiving, by a communication unit, data and a logical address of the data, the data being associated with the ANN, determining, by a processing unit coupled to the communication unit and to a plurality of physical memories and based on the logical address and physical parameters of the physical memories, a physical address of a physical memory of the plurality of physical memories, and performing, by the processing unit, an operation associated with the data and the physical address. The determination of the physical address can be based on a usage count of the data in the ANN computation or a time lapse between a time the data is written to the physical memory and a time the data is used in the ANN computation.

Abstract: Systems and methods for accelerating artificial neural network computation are disclosed. An example may comprise selecting, by a controller communicatively coupled to a selector and an arithmetic unit and based on a criterion, an input value from the stream of input values of a neuron, configuring, by the controller, the selector to provide, dynamically, the selected input value to the arithmetic unit, providing, by the controller to the arithmetic unit, an information of the selected input value, acquiring, by the arithmetic unit and based on the information, a weight from a set of weights, and performing, by the arithmetic unit a mathematical operation on the selected input value and the weight to obtain a result, wherein the result is to be used to compute an output of the neuron. The criterion may include a comparison between the input value and a reference value. The reference value may include zero.

Abstract: Systems and methods for realigning streams of neuron outputs are provided. An example method may include generating, by a processing unit, neuron outputs including at least a first neuron output and a second neuron output, generating, by at least one further processing unit, further neuron outputs including at least a further first neuron output and a further second neuron output, receiving, by a synchronization module communicatively coupled to the processing unit and the further processing unit, the neuron outputs, wherein the neuron outputs and the further neuron outputs are received in an arbitrary order, and ordering, by the synchronization module, the first neuron output, the further first neuron output, the second neuron output and the further second neuron output according to a further order, the further order being different from the arbitrary order.