Brain like large scale neural network system

Continuously optimizing the training data set and neural network model scale is the key to the success of deep learning. When the data set is rich enough, increasing the scale of the neural network can obtain higher prediction accuracy, which has been verified in a series of research fields including text, image, audio, etc. However, the current problems of modular neural network design are mainly concentrated in a single field, solving problems based on data characteristics, rather than discovering the semantic relationship of data like the human brain, so the interpretability and robustness are not good. This project proposes a method to build a super large-scale neural network system based on the functional mechanism of the human brain. The system is based on brain-like and neural network research, integrating a large amount of existing data and pre-training models, and breaking through the existing deep neural network processing functions The single defect aims to design a new generation of neural network models, realize human-like knowledge storage and reasoning, and move towards general artificial intelligence. The entire system includes the foundation of brain function research, the construction of a brain-like functional neural network library, and how to design a large-scale neural network model component platform based on the brain function collaborative work mechanism and the brain-like neural network library. At the same time, in order to achieve efficient training of super-large-scale neural networks, the system provides a distributed neural network training platform and training algorithms. Finally, the system is verified on multiple different modal tasks.