While it’s straightforward to assume that machine learning is an entirely digital affair, born from the intersection of computers and algorithms mimicking cerebral functions.
While early machines were analog in nature, a growing body of research is now suggesting that mechanical systems are capable of learning as well. Physicists at the University of Michigan have made a notable contribution to this field of research.
Researchers Shuaifeng Li and Xiaoming Mao, part of the University of Michigan crew, developed an algorithm providing a mathematical framework for how studies operate within lattice structures known as mechanical neural networks.
As researchers discovered, autonomous systems are capable of receiving instructions and performing calculations independently, a breakthrough that holds significant implications for the future of artificial intelligence.
Researchers have demonstrated the efficacy of this algorithm in “training” supply data to tackle complex problems, such as identifying various species of iris plants. As advancements unfold, self-sustaining structures may emerge, capable of resolving increasingly complex challenges – akin to airplane wings adapting their design to diverse wind conditions – without human intervention or computational assistance.
The potential implications of this research are far-reaching, with its findings potentially sparking innovative ideas not only within the field but also among experts outside it, said Li, a postdoctoral researcher.
The algorithm hinges on the method of backpropagation, a technique that has proven effective in facilitating learning within both digital and optical frameworks. Because of its inherent lack of bias towards data transmission, this algorithm can potentially unlock fresh pathways for exploring how resident programs learn.
“We’re witnessing the widespread application of backpropagation principles in numerous biological systems,” Li noted. “This understanding could also help biologists better grasp the workings of organic neural networks in humans and other species.”
Physicists Li and Mao from the University of Michigan’s Division of Physics have published their groundbreaking study in the esteemed journal Nature Communications.
The idea of harnessing bodily objects in computation has been around for decades. While mechanical neural networks are a relatively recent development, their potential has sparked growing interest as AI continues to evolve and advance rapidly?
Significant advancements in technology have largely taken place in the field of computer science, with numerous groundbreaking innovations being widely observed. Each week, tens of millions of individuals worldwide turn to AI-powered chatbots like ChatGPT for assistance with tasks such as composing emails, planning vacations, and more.
AI assistants primarily rely on artificial neural network architectures. While their inner mechanics may be complex and obscure, likening them to mechanical neural networks provides a valuable framework for understanding the functionality of such systems, according to Li.
As consumers interact with a chatbot, they compose an input command or query, which is parsed by a sophisticated neural network algorithm running on a powerful computer network. Based on its training data, the system produces an output that appears on the user’s screen.
A mechanical neural network, or MNN, shares fundamental components identically. Li and Mao’s research featured a novel approach where a weighted cloth served as an interactive device, effectively engaging with the processing system. As the weight exerted its influence, the fabric underwent a transformation, its shape and structure subtly adjusting to accommodate the pressure.
The concept is that processing power enters data and supplies deform to produce an output or response, much like a processor processes information and generates results.
Researchers employed three-dimensional printing technology to produce customized lattices, comprising interconnected triangular units that were subsequently transformed into larger trapezoidal structures. Supplies can be tailored by modulating the rigidity or pliability of specific components within that framework.
To achieve their intended futuristic functions, micro-nano-networks (MNNs) must be capable of dynamically adjusting these segments as needed, much like an aircraft’s wings modify their characteristics in real-time. Supplies being researched for potential uses, though they cannot currently be categorized by inventory list or database.
Li modelled this habit by iteratively printing out novel processor variations featuring adjusted phase thicknesses, thereby achieving the desired performance output. Li and Mao’s key innovation lies in developing an algorithm that enables a fabric to dynamically adjust its constituent parts.
While the mathematical nuances of backpropagation may be complex, its underlying idea remains straightforward and accessible, according to Li.
To initiate the process effectively, it’s essential to clarify your input parameters and specify how you want the system to respond accordingly? The precise response is sought. As the community leverages this distinction, it adapts its approach to converge towards the target outcome across successive iterations.
Mathematically, the difference between the actual output and the specified output is mathematically described by the concept of a loss function. By employing a mathematical operator known as the gradient to operate on the loss function, the community learns how to adapt and improve.
Li verified that once you’re familiar with the specific information you’re seeking, his Multi-Network Neural Networks (MNNs) effectively display that relevant data.
Li explained that the technology might potentially generate the gradient mechanically, with assistance from cameras and computer code in his study. “It’s surprisingly practical and environmentally conscious.”
Consider a scenario where a lattice entirely comprises uniform elements with identical dimensions and mechanical properties. As the central hub is grasped, its adjacent nodes, situated on either side where branches converge, automatically redistribute equal forces due to the inherent symmetrical design of the system.
Suppose, as an alternative, one sought to craft a lattice yielding the most irregular, unpredictable outcome imaginable. Here is the rewritten text:
To establish a community, it’s essential to draw a clear distinction between nodes situated on the left of a load and those positioned immediately adjacent to it.
Researchers Li and Mao employed their proprietary algorithm in conjunction with a straightforward experimental setup to generate the lattice, yielding the desired result. Like a biological system, the approach primarily focuses on the interactions between proximal entities, paralleling the manner in which neurons operate.
The researchers also provided extensive datasets of input forces, analogous to those used in machine learning on computers, to train their multi-layer neural networks (MNNs).
Notably, distinct instances of environmental stimuli were linked to distinctly varying sizes of petals and leaves on iris plants, underscoring the pivotal role these characteristics play in distinguishing between species. If Li were to present an unknown plant species to the well-versed expert, they would likely be able to accurately classify it.
As Dr. And Li continues to build upon the intricacies of the system, she is poised to tackle its potential applications through Multi-Neural Networks (MNNs) capable of harnessing the power of sound waves.
“We’re capable of encoding significantly more data within the content,” Li said. “With sound waves, one considers amplitude, frequency, and the encoding of information.”
While exploring similar applications, the University of Michigan’s research team is concurrently developing a deeper understanding of network dynamics in supply chains, including polymers and nanoparticle arrangements. Together, they will develop novel software applications where their proprietary algorithm will be implemented, ultimately driving the creation of fully autonomous learning systems.
This research is funded by the Office of Naval Analysis and the National Science Foundation’s Center for Advanced Particle Systems, or COMPASS.
