Reconfigurable pooling operation module structure oriented to neural network and implementation method thereof

Abstract

The invention discloses a reconfigurable pooling operation module structure oriented to a neural network, comprising a dynamic allocation module and a pooling calculation module, wherein the pooling calculation module is used for completing a calculation process of a pooling layer, realizes different pooling methods under the action of the dynamic allocation module and dynamically allocates a calculation scale of a pooling operation; and a current input queue reads from a data storage device after the previous calculation, and a result is transmitted to the data storage device by virtue of anoutput queue. The invention also discloses an implementation method of the reconfigurable pooling operation module structure oriented to the neural network, by virtue of an external control signal, each unit of the dynamic allocation module realizes control on the pooling calculation module, and according to control signal sent out by all the units, the calculation module carries out a corresponding calculation operation and gives an accurate output result. The technical scheme not only can reduce complexity and time of a calculation process and lower power consumption but also improves calculation flexibility and unit utilization rate.

Description

technical field [0001] The invention belongs to the technical field of artificial neural networks, in particular to a reconfigurable pooling operation module structure oriented to neural networks. Background technique [0002] As the network becomes larger and larger, the neural network system needs to perform corresponding calculation operations on a large amount of input data, which will generate a large amount of power consumption. At present, a lot of research has been done on algorithms, software, and hardware. However, low energy consumption still faces new challenges. [0003] By compressing the feature map, the pooling operation can reduce the size of the feature map and simplify the computational complexity of the network; on the other hand, perform feature compression to extract main features. The pooling operation shows superiority in terms of performance and flexibility when dealing with neural networks. [0004] Since the pooling operation is an important proc...

AI Technical Summary

Problems solved by technology

However, low energy consumption still faces new challenges

Images