An On-Chip Memory Compression Method for Neuromorphic Processors for Liquid State Machines

Abstract

The invention discloses a liquid state machine-oriented neuromorphic processor on-chip storage compression method. First, all in-degree weights of neurons are divided into equal groups, and then all non-zero weights are transmitted to corresponding groups in the processor memory. Then use the high-order address to generate a label, distinguish a weight from other weights in the same group, and store it together with all connection information. The on-chip memory is organized as a read-only group associative cache. During the calculation process, according to Request the index of the weight value, generate the group number, access the target group, take out all the labels in the group and compare with the label of the request weight value, the present invention can reduce the power consumption of the processor by utilizing the sparseness of the liquid state machine to compress the storage, It is also possible to make a single processor have more logical neurons to handle more complex tasks without adding additional storage. With different weight data widths, CSSAC can make the processor store up to 14%‑55% reduction and 5%‑46% reduction in power consumption.

Description

technical field [0001] The invention relates to the technical field of a liquid state machine-oriented neuromorphic processor on-chip storage compression method, in particular to an on-chip storage compression method for the weights of connections between neurons in a liquid state machine-oriented neuromorphic processor. Background technique [0002] Spiking neural networks and brain-like processors have received extensive attention and rapid development due to their ability to simulate the behavior of brain neurons and their high energy efficiency. As a kind of spiking neural network, liquid state machine has shown great potential in the fields of image recognition and speech recognition. Because liquid state machines are used to identify the output of various novel sensors, such as dynamic vision sensors (DVS) and dynamic audio sensors (DAS), the resulting pulse train is a good fit. Emerging sensors are capable of capturing dynamic changes within the sensor's field of vie...

AI Technical Summary

Problems solved by technology

However, if the network deployed on a brain-like processor is a sparse liquid state machine, the storage reserved for weights and synapses is redundant

It limits the number of liquid state machine logic neurons that can be supported by a single processor with a fixed area, thus affecting the processing power and optimal accuracy of the liquid state machine

Images