|

Overview

 

​

Brain-inspired Cognitive Intelligence Engine (BrainCog) is a brain-inspired spiking neural network based platform for simulating the cognitive brains of different animal species at multiple scales and realizing brain-inspired Artificial Intelligence. The long term goal of BrainCog is to provide a comprehensive theory and system to decode the mechanisms and principles of human intelligence and its evolution, and develop artificial brains for brain-inspired conscious living machines in future human-machine society.

 

BrainCog integrates neural computational models at different levels of granularity, multiple brain-inspired learning and plasticity principles, and different types of neural network connectivity patterns and encoding strategies.
Towards intelligent models for AI applications, BrainCog realises multiple brain-inspired cognitive functions based on spiking neural networks, which can be classified into five categories: Perception and Learning, Decision Making, Motor Control, Knowledge Representation and Reasoning, and Social Cognition. Together, these components form neural circuits that correspond to 28 brain areas in the mammalian brain.
 

Brain-inspired learning mechanism: We combine local and global plasticity and propose a more biologically plausible spiking neural network with feedforward and feedback connections. The biologically plausible spatio-temporal adaptation algorithm based on BrainCog can theoretically achieve competitive classification accuracy with only about 3% of the energy compared to artificial neural networks of the same structure. The ANN-SNN conversion model built based on BrainCog fully combines the advantages of the backpropagation algorithm and SNN, allowing SNN to perform image classification and target detection tasks almost losslessly with 1/10 and 1/50 of the simulation time of other algorithms. The unsupervised SNN model based on the STDP algorithm supported by BrainCog achieves the best performance among STDP-based unsupervised algorithms to date by integrating a biologically plausible optimisation algorithm and various adaptive mechanisms, while achieving 4-5% better performance than ANN under the same model structure with an extremely small number of samples.
 

Social cognition spiking neural network models enable humanoid robots to perceive and understand themselves and others in order to pass the MultiRobots Mirror Self-Recognition Test, and to help other agents avoid potential risks by showing preliminary ethical-like behaviour.
 

BrainCog provides structural simulations of different mammalian brains at multiple scales, building the mouse brain simulator with different types of punctate neurons, the monkey brain simulator (1.21 billion spiking neurons, 1.3 trillion synapses, 1/5 of the macaque brain scale) and the human brain simulator (860 million spiking neurons, 2.5 trillion synapses, 1/100 of the human brain scale).
 

The BORN, an artificial intelligence engine based on the BrainCog, could perform the emotion-dependent music composition and performance by a humanoid robot.

 

BrainCog aims to provide infrastructural support for specialised, generalised brain-inspired AI. It currently provides cognitive function components that can be grouped into five categories: Perception and Learning, Decision Making, Motor Control, Knowledge Representation and Reasoning, and Social Cognition. Together, these components form neural circuits that correspond to 28 brain areas in the mammalian brain. These brain-inspired AI models have been effectively validated on various supervised and unsupervised learning, deep reinforcement learning, and several complex brain-inspired cognitive tasks.

​

​