INFORMATION TECHNOLOGY

Make AI more biological! Scientists propose new strategies for neural networks


Recently, the brain-like cognitive intelligence team led by Zeng Yi, a researcher at the Institute of Automation of the Chinese Academy of Sciences, published a new study entitled “Brain-inspired neural circuit evolution enabling pulsed neural networks” in the Proceedings of the National Academy of Sciences (PNAS). Inspired by “the diversity of naturally evolved biological brain neural circuit structures” and “pulse timing-dependent plasticity mechanism”, they proposed brain-inspired neural circuit evolution strategies, which are expected to help the industry develop more biologically rational and efficient brain-like pulse neural networks.

In the biological nervous system, different types of neurons can self-organize into neural circuits with different connection patterns to structurally support rich cognitive functions. Different types of neural circuits in the human brain and their adaptive abilities facilitate human perception, learning, decision-making, and other higher cognitive functions. However, most of the current neural network design paradigms are based on structural heuristics in the field of deep learning. These structures are dominated mainly by “feed-forward connections” without taking into account the different types of neurons, which significantly hinders pulsed neural networks from realizing their potential on complex tasks. Therefore, it is still a profound and open challenge to explore the rich dynamics and significance of biological neural circuits from the perspective of computation and apply them to the structure of current brain-like pulse neural networks to improve the capabilities of artificial intelligence (AI) systems.

Based on the combination of “feed-forward” and “feedback” connections with excitatory and inhibitory neurons, Zeng Yi’s team provides a more biologically rational evolutionary space for computational modeling of intelligent evolution. This study uses the local impulse behavior of neurons to adaptively evolve functional neural circuits generated through natural evolution, such as forward excitation, forward inhibition, feedback inhibition, and lateral inhibition, through local rules of pulse timing dependence on plasticity, and update synaptic weights in combination with global error signals. By incorporating evolutionary neural circuits, brain-like pulse neural networks for image classification and reinforcement learning and decision-making tasks were constructed. Using the brain-inspired neural circuit evolution strategy (NeuEvo) and its rich neural circuit types, the evolved brain-like pulsed neural network greatly enhances perception, reinforcement learning and decision-making capabilities.

A brain-like pulse neural network constructed using brain-inspired neural evolution. Image source: Paper

According to reports, NeuEvo has achieved state-of-the-art performance on the CIFAR10, DVS-CIFAR10, DVS-Gesture and N-Caltech101 datasets known at the time of submission, and achieved representative accuracy on pulsed neural networks on ImageNet. Combining online and offline deep reinforcement learning algorithms, it achieves performance comparable to artificial neural networks.

“The evolved brain-like pulsed neural circuits laid the foundation for the evolution of networks with complex functions and the emergence of cognitive abilities.” Zeng Yi said that this study simulates the “use in and waste retreat” in the evolution of natural structures by computational modeling, and independently evolves a variety of neural circuit types on this basis. What’s more interesting is that these circuit types exist in the brains of natural organisms, and experiments have proved that these structures can better help solve core problems related to intelligence such as learning and decision-making, and “existence is reasonable” in natural evolution, which will inspire future research on general-brain-like cognitive intelligence. (Source: China Science News Zhao Guangli)

Related paper information:https://www.pnas.org/doi/10.1073/pnas.2218173120



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button