A method for neural network computation using adaptive data representation, adapted for a processor to perform multiply-and-accumulate operations on a memory having a crossbar architecture, is provided. The memory comprises multiple input and output lines crossing each other, multiple cells respectively disposed at intersections of the input and output lines, and multiple sense amplifiers respectively connected to the output lines. In the method, an input cycle of kth bits respectively in an input data is adaptively divided into multiple sub-cycles, wherein a number of the divided sub-cycles is determined according to a value of k. The kth bits of the input data are inputted to the input lines with the sub-cycles and computation results of the output lines are sensed by the sense amplifiers. The computation results sensed in each sub-cycle are combined to obtain the output data corresponding to the kth bits of the input data.

Apparatuses and methods for implementing artificial synapses utilizing SSM cells. A leaky-integrate-and-fire circuit can provide a feedback signal to an SSM cell responsive to a threshold quantity of pulses being applied to the gate from the signal line. A resulting state of the SSM cell can be dependent on a time difference between a latest of the threshold quantity of pulses and an initial pulse of the feedback signal.

Methods and apparatus are disclosed for managing the storage of static and dynamic neural network data within a non-volatile memory (NVM) die for use with deep neural networks (DNN). Some aspects relate to separate trim sets for separately configuring a static data NVM array for static input data and a dynamic data NVM array for dynamic synaptic weight data. For example, the static data NVM array may be configured via one trim set for data retention, whereas the dynamic data NVM array may be configured via another trim set for write performance. The trim sets may specify different configurations for error correction coding, write verification, and read threshold calibration, as well as different read/write voltage thresholds. In some examples, neural network regularization is provided within a DNN by setting trim parameters to encourage bit flips to avoid overfitting. Some examples relate to managing non-DNN data, such as stochastic gradient data.

Methods and apparatus are disclosed for managing the storage of static and dynamic neural network data within a non-volatile memory (NVM) die for use with deep neural networks (DNN). Some aspects relate to separate trim sets for separately configuring a static data NVM array for static input data and a dynamic data NVM array for dynamic synaptic weight data. For example, the static data NVM array may be configured via one trim set for data retention, whereas the dynamic data NVM array may be configured via another trim set for write performance. The trim sets may specify different configurations for error correction coding, write verification, and read threshold calibration, as well as different read/write voltage thresholds. In some examples, neural network regularization is provided within a DNN by setting trim parameters to encourage bit flips to avoid overfitting. Some examples relate to managing non-DNN data, such as stochastic gradient data.

In a reservoir element that has a plurality of vibrators, at least one of the plurality of vibrators has a vibration state that is different from the vibration states of the other vibrators. The vibrations of the plurality of vibrators are configured to affect each other.

A method and apparatus with artificial network generation and/or implementation corresponding to a natural neural network are disclosed. The method includes performing an adjusting, based on a firing time difference between an action potential (AP) of a first biological neuron and a post-synaptic potential (PSP) of a second biological neuron, of a first conductance value of a first memory corresponding to the first and second biological neurons, adjusting the firing time difference based on the PSP of the second biological neuron, and performing another adjusting, by controlling another firing time difference between the AP of the first biological neuron and the PSP of the second biological neuron to be the adjusted firing time difference, of a second conductance value of the first memory corresponding to the first and second biological neurons.

Provided are an integrated circuit chip apparatus and a related product, the integrated circuit chip apparatus being used for executing a multiplication operation, a convolution operation or a training operation of a neural network. The present technical solution has the advantages of a small amount of calculation and low power consumption.

Provided are an integrated circuit chip apparatus and a related product, the integrated circuit chip apparatus being used for executing a multiplication operation, a convolution operation or a training operation of a neural network. The present technical solution has the advantages of a small amount of calculation and low power consumption.

A method includes receiving one or more input Artificial Intelligence (AI) networks, transforming the AI networks into respective graphs including interconnected logical operators, and mapping the graphs onto a design of a hardware accelerator including a plurality interconnected hardware engines. A performance of running the AI networks on the design of the hardware accelerator is simulated using a petri-net simulation.

A method includes receiving one or more input Artificial Intelligence (AI) networks, transforming the AI networks into respective graphs including interconnected logical operators, and mapping the graphs onto a design of a hardware accelerator including a plurality interconnected hardware engines. A performance of running the AI networks on the design of the hardware accelerator is simulated using a petri-net simulation.