Reconfigurable autonomous learning spiking neural network processor

Abstract

The invention provides a reconfigurable autonomous learning spiking neural network processor. The reconfigurable autonomous learning spiking neural network processor comprises a processing unit arraycomposed of a plurality of processing units and channels in the east, south, west and north directions. Each processing unit comprises an external routing module, a sequencing module, a pulse queue module, a controller module, a search module, a memory module, a client module, a server module, an internal routing module, a plurality of exclusive computing resources and a plurality of borrowable computing resources. The processor adopts a pulse packet composed of pulse generation time and a source neuron ID to transmit signals; the computing executing modes of the computing resources include aninference mode and a learning mode, and a target neuron membrane potential, a synaptic weight related variable or a synaptic weight is updated by adopting a reconfigurable circuit; the computing resources comprise a self-adaptive clock-driven and event-driven computing mechanism module, and according to the updating time interval, the computing mode that the computing unit executes updating computing is changed in a self-adaptive mode.

Description

technical field [0001] The invention relates to the technical field of pulse neural network, in particular to a reconfigurable self-learning pulse neural network processor. Background technique [0002] Neural network is a mathematical model abstracted according to the characteristics of biological brain nervous system, which can be used to solve practical problems such as object recognition. Its development process can be divided into three generations: Perceptron, Artificial Neural Network (ANN), Pulse Neural Network (Spiking Neural Network, SNN). Early perceptrons could only solve linear problems, but were powerless to nonlinear problems (such as XOR problems); ANN added a nonlinear activation function on the basis of perceptrons to enable it to solve nonlinear problems. In recent years, with the improvement of computer computing power, ANN has a better performance in solving some complex problems, but with the increase of the complexity of the problem, a deeper network ...

AI Technical Summary

Problems solved by technology

Due to the relatively large network structure, this design requires high transmission bandwidth of the system, resulting in reduced system performance

[0006] 3. In order to support the online learning of SNN, the existing design needs to design an additional learning circuit, which greatly increases the overhead of hardware resources

Therefore, for different pulse input situations, different update methods should be used in order to make the power consumption of the system as low as possible, but the current design does not take this into consideration.

Images