Abstract:Neural network is one of the mainstream interests in artificial intelligence research. However, neurons in most artificial neural networks are in a fixed connection with each other at this stage, making it impossible to achieve self-learning ability. In this paper, we present our approach to build a self-growing and self-organizing neural network. We try to imitate the way in biological neural network, where neurons would form connections as a response to external stimulates, and evolve, thus to build a self-growing, self-organizing neural network. The neuron model in our work is built based on Leaky I&F model [1]. The inputs of our model are digital pulse signals. A counter was added into the neuron model, so that it could calculate the changing rate of outputs, called the output intensity of a neuron. Only when the output intensity of a neuron meets a certain condition, it can exchange information with other neurons. The distance of the forming connection is decided by the output intensity. While the target of connection is picked through calculation of a Probability Matrix C, the values in which are generated by training. Matrix C could learn from the input data, thus to optimize the generation of neural networks. We verify the generated neural network with different inputs. Those could generate different outputs with respect to different inputs would be treated as successful ones. Redundant connections is then deleted. We have exploited a CPU+GPU way to build the simulation environment. CPU is used to control, while GPU is used for all the computation of neurons. 2-bits input is first used to stimulate the network, and prompted it to grow. After verification and optimization, those generated seeds are used to build more complex ones. The result shows that, neural network built by our method can self-grow and selforganize as the complexity of the input external signals increase. Also our neural network can show some level of intelligence.

Mechanical Neural Networks with Explicit and Robust Neurons

Training all-mechanical neural networks for task learning through in situ backpropagation

Neuromechanical Autoencoders: Learning to Couple Elastic and Neural Network Nonlinearity

In-memory Mechanical Computing

Intelligent mechanical metamaterials towards learning static and dynamic behaviors

Mechanistic Neural Networks for Scientific Machine Learning

Hybrid Dual-Complementary Metal-Oxide-semiconductor/memristor Synapse-Based Neural Network with Its Applications in Image Super-Resolution

Novel neural controller for robot manipulator

Automatic Structure and Parameter Training Methods for Modeling of Mechanical System by Recurrent Neural Networks

Mechanical Cognitivization: a Kinematic System Proof of Concept.

Design of Artificial Neurons of Memristive Neuromorphic Networks Based on Biological Neural Dynamics and Structures

Metamaterial-Based Analog Recurrent Neural Network Toward Machine

Metamaterial-Based Analog Recurrent Neural Network Toward Machine Intelligence

Spiking recurrent neural networks for neuromorphic computing in nonlinear structural mechanics

In Situ Training of Feed-Forward and Recurrent Convolutional Memristor Networks

Physical Neural Networks with Self-Learning Capabilities

Bioinspired Gain-Modulated Recurrent Neural Network for Controlling Musculoskeletal Robot

Dynamic neural network with memristive CIM and CAM for 2D and 3D vision

Complex Learning in Bio-plausible Memristive Networks

Dynamic neuronal networks efficiently achieve classification in robotic interactions with real-world objects

Exploration Of A Mechanism To Form Bionic, Self-Growing And Self-Organizing Neural Network