Abstract: Embodiments of the invention provide a method comprising creating a structural description for at least one neurosynaptic core circuit. Each core circuit comprises an interconnect network including plural electronic synapses for interconnecting one or more electronic neurons with one or more electronic axons. The structural description defines a desired neuronal activity for the core circuits. The desired neuronal activity is simulated by programming the core circuits with the structural description. The structural description controls routing of neuronal firing events for the core circuits.

Abstract: Neural inference chips for computing neural activations are provided. In various embodiments, a neural inference chip comprises at least one neural core, a memory array, an instruction buffer, and an instruction memory. The instruction buffer has a position corresponding to each of a plurality of elements of the memory array. The instruction memory provides at least one instruction to the instruction buffer. The instruction buffer advances the at least one instruction between positions in the instruction buffer. The instruction buffer provides the at least one instruction to at least one of the plurality of elements of the memory array from its associated position in the instruction buffer when the memory of the at least one of the plurality of elements contains data associated with the at least one instruction. Each element of the memory array provides a data block from its memory to its horizontal buffer in response to the arrival of an associated instruction from the instruction buffer.

Abstract: Embodiments of the invention relate to canonical spiking neurons for spatiotemporal associative memory. An aspect of the invention provides a spatiotemporal associative memory including a plurality of electronic neurons having a layered neural net relationship with directional synaptic connectivity. The plurality of electronic neurons configured to detect the presence of a spatiotemporal pattern in a real-time data stream, and extract the spatiotemporal pattern. The plurality of electronic neurons are further configured to, based on learning rules, store the spatiotemporal pattern in the plurality of electronic neurons, and upon being presented with a version of the spatiotemporal pattern, retrieve the stored spatiotemporal pattern.

Abstract: Embodiments of the invention provide a method comprising receiving a set of features extracted from input data, training a linear classifier based on the set of features extracted, and generating a first matrix using the linear classifier. The first matrix includes multiple dimensions. Each dimension includes multiple elements. Elements of a first dimension correspond to the set of features extracted. Elements of a second dimension correspond to a set of classification labels. The elements of the second dimension are arranged based on one or more synaptic weight arrangements. Each synaptic weight arrangement represents effective synaptic strengths for a classification label of the set of classification labels. The neurosynaptic core circuit is programmed with synaptic connectivity information based on the synaptic weight arrangements. The core circuit is configured to classify one or more objects of interest in the input data.

Abstract: Learned step size quantization in artificial neural network is provided. In various embodiments, a system comprises an artificial neural network and a computing node. The artificial neural network comprises: a quantizer having a configurable step size, the quantizer adapted to receive a plurality of input values and quantize the plurality of input values according to the configurable step size to produce a plurality of quantized input values, at least one matrix multiplier configured to receive the plurality of quantized input values from the quantizer and to apply a plurality of weights to the quantized input values to determine a plurality of output values having a first precision, and a multiplier configured to scale the output values to a second precision.

Abstract: In one embodiment, the present invention provides a neural network comprising multiple modalities. Each modality comprises multiple neurons. The neural network further comprises an interconnection lattice for cross-associating signaling between the neurons in different modalities. The interconnection lattice includes a plurality of perception neuron populations along a number of bottom-up signaling pathways, and a plurality of action neuron populations along a number of top-down signaling pathways. Each perception neuron along a bottom-up signaling pathway has a corresponding action neuron along a reciprocal top-down signaling pathway. An input neuron population configured to receive sensory input drives perception neurons along a number of bottom-up signaling pathways. A first set of perception neurons along bottom-up signaling pathways drive a first set of action neurons along top-down signaling pathways.

Abstract: Embodiments of the invention relate to providing transposable access to a synapse array using a recursive array layout. One embodiment comprises maintaining synaptic weights for multiple synapses connecting multiple axons and multiple neurons, wherein the synaptic weights are maintained based on a recursive array layout. The recursive array layout facilitates transposable access to the synaptic weights. A neuronal spike event between an axon and a neuron is communicated via a corresponding connecting synapse by accessing the synaptic weight of the corresponding connecting synapse in the recursive array layout.

Abstract: A neural inference chip includes a global weight memory; a neural core; and a network connecting the global weight memory to the at least one neural core. The neural core comprises a local weight memory. The local weight memory comprises a plurality of memory banks. Each of the plurality of memory banks is uniquely addressable by at least one index. The neural inference chip is adapted to store in the global weight memory a compressed weight block comprising at least one compressed weight matrix. The neural inference chip is adapted to transmit the compressed weight block from the global weight memory to the core via the network. The core is adapted to decode the at least one compressed weight matrix into a decoded weight matrix and store the decoded weight matrix in its local weight memory. The at core is adapted to apply the decoded weight matrix to a plurality of input activations to produce a plurality of output activations.

Abstract: In one embodiment, the present invention provides a neural network circuit comprising multiple symmetric core circuits. Each symmetric core circuit comprises a first core module and a second core module. Each core module comprises a plurality of electronic neurons, a plurality of electronic axons, and an interconnection network comprising multiple electronic synapses interconnecting the axons to the neurons. Each synapse interconnects an axon to a neuron. The first core module and the second core module are logically overlayed on one another such that neurons in the first core module are proximal to axons in the second core module, and axons in the first core module are proximal to neurons in the second core module. Each neuron in each core module receives axonal firing events via interconnected axons and generates a neuronal firing event according to a neuronal activation function.

Abstract: The present invention provides a system comprising multiple core circuits. Each core circuit comprises multiple electronic axons for receiving event packets, multiple electronic neurons for generating event packets, and a fanout crossbar including multiple electronic synapse devices for interconnecting the neurons with the axons. The system further comprises a routing system for routing event packets between the core circuits. The routing system virtually connects each neuron with one or more programmable target axons for the neuron by routing each event packet generated by the neuron to the target axons. Each target axon for each neuron of each core circuit is an axon located on the same core circuit as, or a different core circuit than, the neuron.

Abstract: Neural inference chips are provided. A neural core of the neural inference chip comprises a vector-matrix multiplier; a vector processor; and an activation unit operatively coupled to the vector processor. The vector-matrix multiplier, vector processor, and/or activation unit is adapted to operate at variable precision.

Abstract: The present invention provides a system comprising multiple core circuits. Each core circuit comprises multiple electronic axons for receiving event packets, multiple electronic neurons for generating event packets, and a fanout crossbar including multiple electronic synapse devices for interconnecting the neurons with the axons. The system further comprises a routing system for routing event packets between the core circuits. The routing system virtually connects each neuron with one or more programmable target axons for the neuron by routing each event packet generated by the neuron to the target axons. Each target axon for each neuron of each core circuit is an axon located on the same core circuit as, or a different core circuit than, the neuron.

Abstract: Massively parallel neural inference computing elements are provided. A plurality of multipliers is arranged in a plurality of equal-sized groups. Each of the plurality of multipliers is adapted to, in parallel, apply a weight to an input activation to generate an output. A plurality of adders is operatively coupled to one of the groups of multipliers. Each of the plurality of adders is adapted to, in parallel, add the outputs of the multipliers within its associated group to generate a partial sum. A plurality of function blocks is operatively coupled to one of the plurality of adders. Each of the plurality of function blocks is adapted to, in parallel, apply a function to the partial sum of its associated adder to generate an output value.

Abstract: Three-dimensional neural inference processing units are provided. A first tier comprises a plurality of neural cores. Each core comprises a neural computation unit. The neural computation unit is adapted to apply a plurality of synaptic weights to a plurality of input activations to produce a plurality of output activations. A second tier comprises a first neural network model memory adapted to store the plurality of synaptic weights. A communication network is operatively coupled to the first neural network model memory and to each of the plurality of neural cores, and adapted to provide the synaptic weights from the first neural network model memory to each of the plurality of neural cores.

Abstract: A multiplexed neural core circuit according to one embodiment comprises, for an integer multiplexing factor T that is greater than zero, T sets of electronic neurons, T sets of electronic axons, where each set of the T sets of electronic axons corresponds to one of the T sets of electronic neurons, and a synaptic interconnection network comprising a plurality of electronic synapses that each interconnect a single electronic axon to a single electronic neuron, where the interconnection network interconnects each set of the T sets of electronic axons to its corresponding set of electronic neurons.

Abstract: Embodiments of the invention provide a method for mapping a bipartite graph onto a neuromorphic architecture comprising of a plurality of interconnected neuromorphic core circuits. The graph includes a set of source nodes and a set of target nodes. The method comprises, for each source node, creating a corresponding splitter construct configured to duplicate input. Each splitter construct comprises a first portion of a core circuit. The method further comprises, for each target node, creating a corresponding merger construct configured to combine input. Each merger construct comprises a second portion of a core circuit. Source nodes and target nodes are connected based on a permutation of an interconnect network interconnecting the core circuits.

Abstract: Embodiments of the invention provide a system and circuit interconnecting peripheral devices to neurosynaptic core circuits. The neurosynaptic system includes an interconnect that includes different types of communication channels. A device connects to the neurosynaptic system via the interconnect.

Abstract: Neural inference chips for computing neural activations are provided. In various embodiments, the neural inference chip is adapted to: receive an input activation tensor comprising a plurality of input activations; receive a weight tensor comprising a plurality of weights; Booth recode each of the plurality of weights into a plurality of Booth-coded weights, each Booth coded value having an order; multiply the input activation tensor by the Booth coded weights, yielding a plurality of results for each input activation, each of the plurality of results corresponding to the orders of the Booth-coded weights; for each order of the Booth-coded weights, sum the corresponding results, yielding a plurality of partial sums, one for each order; and compute a neural activation from a sum of the plurality of partial sums.

Abstract: One embodiment provides a system comprising a memory device for maintaining deterministic neural data relating to a digital neuron and a logic circuit for deterministic neural computation and stochastic neural computation. Deterministic neural computation comprises processing a neuronal state of the neuron based on the deterministic neural data maintained. Stochastic neural computation comprises generating stochastic neural data relating to the neuron and processing the neuronal state of the neuron based on the stochastic neural data generated.

Abstract: The present invention provides an event-driven universal neural network circuit. The circuit comprises a plurality of neural modules. Each neural module comprises multiple digital neurons such that each neuron in a neural module has a corresponding neuron in another neural module. An interconnection network comprising a plurality of digital synapses interconnects the neural modules. Each synapse interconnects a first neural module to a second neural module by interconnecting a neuron in the first neural module to a corresponding neuron in the second neural module. Corresponding neurons in the first neural module and the second neural module communicate via the synapses. Each synapse comprises a learning rule associating a neuron in the first neural module with a corresponding neuron in the second neural module. A control module generates signals which define a set of time steps for event-driven operation of the neurons and event communication via the interconnection network.