Over the past year, generative AI models like ChatGPT and DALL-E have produced massive amounts of high-quality, human-like creative output from basic instructions.
Despite beating humans in big-data pattern identification, present AI systems are not intelligent. AIs don’t learn like humans.
Training AI systems requires much more energy and resources than our three meals a day. They lack human-like memory and cannot adapt to dynamic, unpredictable, and loud situations.
We study non-biological brain-like systems. In a Science Advances study, self-organizing networks of tiny silver wires learned and remembered like our brains.
Neuromorphics research replicates biological neurons and synapses in non-biological devices.
Our research uses “nanowires” to replicate brain neurons and synapses.
Nanowires are one-thousandth the breadth of a human hair. Silver, a highly conductive metal, is coated with plastic.
Nanowires self-assemble into neuronal networks. Each metal nanowire has a tiny insulating layer like neurons.
Like neurotransmitters across synapses, ions travel over the insulating layer and into nearby nanowires when electrical signals excite nanowires. Nanowire networks exhibit synapse-like electrical signaling.
Memory and learning
Our new nanowire system investigates human-like intelligence. Learning and memory are central to our study.
We found that nanowire networks can selectively strengthen and weaken synaptic circuits. This resembles brain “supervised learning”.
This compares synaptic output to a desired result. Next, the synapses are reinforced or pruned depending on their output.
We demonstrated that “rewarding” or “punishing” the network increased strengthening. Brain “reinforcement learning” inspired this.
We implemented the “n-back task” to measure human working memory. It involves providing a sequence of stimuli and comparing each new entry to one from n steps ago.
The network “remembered” signals for seven steps. Curiously, seven is the average number of items humans can maintain in working memory.
Reinforcement learning greatly improved the network’s memory.
In our nanowire networks, synaptic pathways emerge based on past activation. Neuroscientists call this “metaplasticity” for brain synapses.
Replicating human intelligence is unlikely.
Our study on neuromorphic nanowire networks reveals that non-biological hardware may incorporate intelligence-related properties like learning and memory.
Nanowire networks differ from AI neural networks. “Synthetic intelligence” may result.
A neuromorphic nanowire network may learn to have more human-like conversations than ChatGPT and remember them.