How Nanoparticles Form Adaptive Neural Network Connections

This patent describes how to build and strengthen a physical neural network using tiny nanoparticles suspended in a liquid, where electric fields make the connections learn and adapt.

What it covers

This patent outlines methods for creating and modifying a physical neural network, called an electromechanical neural network. It uses many tiny nanoparticles floating in a non-conductive liquid, positioned between 'pre-synaptic' and 'post-synaptic' electrodes (Claim 1). When one neuron fires, it increases the voltage at its pre-synaptic electrode. A special 'refractory pulse' then decreases the voltage at the post-synaptic electrode, increasing the voltage difference across the connection (Claim 1). This increased electric field, especially when applied at a higher frequency, causes the nanoparticles to align better, making the connection stronger and reducing its electrical resistance (Claim 2). The network can adapt by summing signals, comparing them to a threshold, and then grounding the post-synaptic junction during a pulse to further increase the local electric field and strengthen the active connections (Claim 4).

What it doesn't cover

• —Does not cover neural networks implemented purely in software on traditional digital processors, as it specifies 'electromechanical' components and 'nanoconnections'.
• —Does not cover physical neural networks that use different materials or mechanisms to form and strengthen connections, such as optical switches or fixed-wire connections without nanoparticles.
• —Does not cover adaptive neural networks where connection strength is modified without an increase in the local electric field or the physical alignment of nanoparticles.
• —Does not cover neural networks that do not utilize a 'refractory pulse' mechanism to specifically adjust synaptic strength as described in Claim 1.
• —Does not cover networks where the connections are not formed from nanoparticles disposed within a dielectric solution.

The clever bit

The truly novel aspect is using nanoparticles suspended in a liquid to form dynamic, self-organizing connections (nanoconnections) between electrodes. These connections can be strengthened by simply applying electric fields, causing the nanoparticles to align and reduce resistance, effectively 'learning' in a physical way.

Why it matters

This patent is foundational for building hardware that mimics the brain's structure, known as neuromorphic computing. Such physical neural networks offer the potential for significantly more energy-efficient and faster artificial intelligence compared to software-based AI running on traditional computer chips. It addresses the challenge of creating physical, adaptive connections that can learn and change, much like biological synapses.

Real-world examples

• 1.Knowmtech's memristor-based AI hardware
• 2.Experimental neuromorphic computing chips
• 3.AI accelerators designed for energy efficiency