# Advancements in Electrolytic Neural Networks and Memristors
Written on
Chapter 1: Understanding Memristors and Neural Networks
Recent innovations in solid-state and discrete neural networks, particularly utilizing solid-state memristors, have generated significant excitement. A 2023 publication detailed a design leveraging silver nanowire memristor networks, which intriguingly aligns with my own experiments centered around electrolytic neural networks developed from the stripping and plating of copper on resistive materials. This could pave the way for constructing neural networks right in home laboratories.
Memristors are an intriguing component of electronics, differing from standard resistors. While resistors consistently limit power flow, memristors adapt their resistance based on the power passing through them. Some memristors increase resistance, gradually restricting flow, while others decrease it, allowing for a greater flow of energy. This technology is relatively new, with substantial research emerging over the last few decades.
Section 1.1: The Role of Neural Networks
Neural networks have been utilized in both hardware and software for over eight decades, yet they have recently surged in popularity, particularly with the rise of generative AI and large language models such as ChatGPT. Within a neural network, a neuron gathers inputs from other neurons and, upon reaching a certain threshold, activates to transmit values to the subsequent layer. This process operates via a sharp delineation known as a “selection function”—insufficient input results in no output.
Subsection 1.1.1: Memristors and Neural Networks
In both solid-state and electrolytic memristors, there are specific voltages or processes that determine this “selection” threshold. For instance, with copper, if the voltage dips below 0.34V, it tends to plate onto the surface. When a voltage is applied to a resistive surface immersed in copper electrolyte, a voltage gradient emerges across the resistor. Parts of the resistor below 0.34V will deposit copper, thereby altering the surface's overall resistance. This mechanism was explored in my initial article on an experimental open-source memristor and programmable diode.
Section 1.2: Insights from Silver Nanowire Networks
Recent research published in AAAS Science Advances (Neuromorphic learning, working memory, and metaplasticity in nanowire networks) reveals analogous capabilities in silver nanowire solid-state networks. In such networks, semi-conductive layers dynamically change as oxygen atoms shift and rearrange at junctions. These random wires create connection points resembling the "cat's whisker" rectifier diodes from the 1940s.
Chapter 2: Signal Processing and Its Implications
The diagram includes an ADC (Analog to Digital Converter), highlighting the necessity of converting tiny signals into usable outputs. In my setup, each output must be amplified—possibly through a single transistor—to revert to the source voltage, or else the signal diminishes with each layer. Inputs also follow this conversion, where 1s represent 1V and 0s signify “no voltage.”
The first video titled "How to Use a TENS Unit for Sciatic Nerve Pain Relief" explains the utility of TENS units in alleviating sciatic nerve discomfort.
Section 2.1: Effects on Copper-Based Networks
This design parallels my own with the resistive sheet, though with a few distinctions. One critical aspect is the presence of holes, which allow signals to interact more intricately. In a unified sheet, the signals have limited pathways.
The second distinction involves inputs. In my configuration, I envisioned a ground (GND) connection or a negative voltage for “0” values. Their model, however, utilized a “high impedance”—essentially a large resistor—meaning voltage-wise, it’s not a straightforward 1V versus 0V or 1V versus -1V scenario. Only the 1V wires are active, while the rest remain inactive. This subtle alteration will also be integrated into my future experiments.
Exciting Possibilities Ahead
I’m eager to apply these techniques in my home lab and hope to see fellow experimenters creating larger and more complex networks using copper “needles” in the near future.
The next five years hold immense potential for neural networks to revolutionize data processing, generative AI, and large language models. Envision a solid-state AI “brain” that could be utilized for various tasks or even to create companions. These advanced systems might aid in scientific research that is too tedious or hazardous for human scientists and could explore the solar system autonomously for extended periods.
Or perhaps they could enhance video games; I’m open to either possibility.
The second video titled "02252024 Charbone deepdive" provides a thorough exploration of the Charbone initiative and its implications for future technologies.
Additional Insights: Hall Effect
Both solid-state and electrolytic methods incorporate semiconductive materials. By generating a magnetic field within the spaces between nanowires or holes, it may be possible to fine-tune the network during its training phase. Tiny electromagnets could enforce specific pathways, guiding the flow of electrons more intentionally.
Imagine a loop of semiconductive material with wires on either side—under normal conditions, electrons would traverse both sides. However, in the presence of a magnetic field, electrons will favor one path over the other, influenced by the field's polarity. This phenomenon, known as the Hall Effect, is utilized in various applications, including magnetic field sensing. The altered electron flow could impact memristive changes, leading to one path becoming more resistive while another becomes less so. For further reading, check out my article on creating memory using the Hall Effect, titled "One Little Bit: Hall-Effect Core Memory Circuits That Remember Their State."
Citation:
Alon Loeffler et al., Neuromorphic learning, working memory, and metaplasticity in nanowire networks. Sci. Adv. 9, eadg3289 (2023). DOI:10.1126/sciadv.adg3289