Over the past year or so, generative AI models like ChatGPT and DALL-E have made it possible to produce massive amounts of seemingly human-like, high-quality creative content from a simple series of prompts.
While they are highly capable – far outperforming humans in big data pattern recognition tasks in particular – today’s AI systems are not intelligent in the same way that we are. AI systems are not structured like our brains and do not learn in the same way.
AI systems also use stretched out amounts of energy and resources for exercise (compared to our three or so meals a day). Their ability to adapt and function in dynamic, unpredictable, and noisy environments is poor compared to ours, and they lack human-like memory.
Our research explores non-biological systems that are more like human brains. In a new study published in Science Advances, we found that self-organizing networks of tiny silver threads seem to learn and remember in much the same way as the thinking machines in our heads.
Imitating the brain
Our work is part of a field of research called neuromorphics, which aims to mimic the structure and functionality of biological neurons and synapses in non-biological systems.
Our research focuses on a system that uses a network of “nanowires” to mimic the neurons and synapses in the brain. These nanowires are tiny threads about one-thousandth the width of a human hair. They are made of a highly conductive metal, such as silver, usually coated with an insulating material such as plastic.
Nanowires self-assemble to form a network structure similar to a biological neural network. Like neurons, which have an insulating membrane, each metal nanowire is covered with a thin insulating layer.
When we stimulate nanowires with electrical signals, ions migrate across the insulating layer to a neighboring nanowire (similar to neurotransmitters across synapses). As a result, we observe synapse-like electrical signaling in nanowire networks.
Learning and memory
Our new work uses this nanowire system to explore the question of human-like intelligence. At the center of our research are two features indicative of high-order cognitive function: learning and memory.
Our study shows that we can selectively strengthen (and weaken) synaptic pathways in nanowire networks. This is similar to “supervised learningin the brain. In this process, the output of synapses is compared to a desired outcome. Then the synapses are strengthened (if their output is close to the desired result) or pruned (if their output is not close to the desired result).
We extended this result by showing that we could increase the degree of reinforcement by “rewarding” or “punishing” the network. This process was inspired by “reinforcement learningin the brain.
Read more: Neuron-like circuits bring brain-like computers one step closer
We also implemented a version of a test called “N– return taskwhich is used to measure working memory in humans. It involves presenting a series of stimuli and comparing each new item with an item that has gone through a number of steps (N) past.
The network “remembered” previous signals for at least seven steps. Oddly enough, seven is often considered the average number of articles people can stay in working memory all at once.
When we used reinforcement learning, we saw dramatic improvements in network memory performance.
In our nanowire networks, we found that the formation of synaptic pathways depends on how those synapses have been activated in the past. This is also the case for synapses in the brain, where neuroscientists “metaplasticity”.
Human intelligence is probably far from being replicated.
Nevertheless, our research on neuromorphic nanowire networks shows that it is possible to implement functions essential for intelligence, such as learning and memory, in non-biological, physical hardware.
Read more: Five ways the superintelligence revolution could happen
Nanowire networks are different from the artificial neural networks used in AI. Yet they can lead to so-called “synthetic intelligence”.
Perhaps a neuromorphic nanowire network could one day learn to make and remember conversations that are more human than ChatGPT.