Neural network three-body model based on mesoscopic system

Abstract

The invention provides a neural network three-body model based on mesoscopic systems. The neural network three-body model comprises a learning body which is mainly used for construction from original data to a high-level mesoscopic system; a memory body which is mainly used for storing and recombining neuron attribute information in the different mesoscopic systems; and an interpretation body which is mainly used as a reverse propagation process of the learning body so as to realize a corresponding interpretation function of the high-level mesoscopic system to a low-level mesoscopic system. According to the invention, important nodes and intrinsic information in a deep learning process are stored by constructing the three-body neural network model based on the mesoscopic systems, and a memory organization structure is further formed, so that important basis and guarantee are provided for interpretability of deep learning, and a new direction is provided for the development of a brand new neural network which better conforms to a human learning cognition mode.

Description

technical field [0001] The invention belongs to the field of neural network three-body models, in particular to a neural network three-body model based on a mesoscopic system. Background technique [0002] In recent years, in many fields such as industrial inspection and intelligent body operation, the demand for large-scene laser three-dimensional (Three Dimensional, 3D) point cloud target recognition and tracking technology based on deep learning has become increasingly strong. However, due to the irregularity, unstructured, and disordered characteristics of laser point clouds, traditional convolutional neural networks (CNN) cannot be directly applied to such data. [0003] Therefore, in order to meet the needs of complex systems, a mesoscopic system-based neural network three-body model is proposed to store important nodes and intrinsic information in the deep learning process, and then form a memory organization structure, which is an interpretable model for deep learnin...

AI Technical Summary

Problems solved by technology

However, due to the irregularity, unstructured, and disordered characteristics of laser point clouds, traditional Convolutional Neural Networks (CNN) cannot be directly applied to such data.

Images