Mechanical Neural Networks with Explicit and Robust Neurons.

Area of Science:

• Robotics and Artificial Intelligence
• Mechanical Engineering
• Computational Science

Background:

• Mechanical computing integrates sensing, analyzing, and actuation for mechanical intelligence.
• Training mechanical neural networks (MNNs) is computationally intensive due to solving nonlinear equilibrium equations.
• Existing MNNs face challenges in efficiency and robustness for complex cognitive tasks.

Purpose of the Study:

• To develop an explicit mechanical neuron that bypasses the need to solve equilibrium equations.
• To introduce a robust training method for mechanical neurons, ensuring insensitivity to defects and perturbations.
• To demonstrate the assembly and application of explicit and robust mechanical neurons in advanced network structures.

Main Methods:

• Development of an explicit mechanical neuron model.
• Implementation of a novel training methodology focusing on robustness.
• Experimental demonstration of a robust mechanical convolutional neural network (MNN).
• Experimental demonstration of a mechanical recurrent neural network (MNN) with long short-term memory (LSTM) capabilities.

Main Results:

• The explicit mechanical neuron allows direct response determination, eliminating complex equation solving.
• The proposed training method ensures neuron robustness against defects and perturbations.
• Demonstrated successful assembly of various network structures using the new neurons.
• Successfully implemented MNNs for tasks including associative learning.

Conclusions:

• The explicit and robust mechanical neuron significantly streamlines the design and training of MNNs.
• This advancement facilitates the creation of intelligent robotic matter.
• The developed neurons offer a more efficient and reliable approach to mechanical intelligence.