In the past year or two, generative AI models like ChatGPT and DALL-E have
made it feasible to generate enormous amounts of high-quality, creative
content that appears to be created by humans from a limited set of
prompts.
Current AI systems are not as intelligent as humans are, despite being
quite powerful and consistently surpassing people in large data pattern
recognition jobs in particular. AI systems don't learn in the same manner as
human brains and aren't structured similarly.
In comparison to our three or so meals a day, AI systems similarly use
enormous quantities of energy and resources throughout their training phase.
In contrast to humans, they are less able to adapt to and function in
dynamic, unpredictable, and loud settings, and they do not have human-like
memory capacities.
Our study investigates non-biological systems better like human brains. In
a
recent study, which was published in Science Advances, we discovered that
self-organizing networks of minute silver wires seem to learn and remember
in a manner similar to that of the thinking machinery in human brains.
Copying the brain
Our study falls under the umbrella of the discipline known as
neuromorphics, which tries to mimic the structure and operation of
biological neurons and synapses in artificial systems.
Our work focuses on a technology that mimics the brain's synapses and
neurons using a network of "nanowires".
The width of these nanowires, which are very small, is equivalent to that
of a human hair. They are often covered in an insulating substance like
plastic and constructed of a highly conductive metal like silver.
.webp)
Self-assembling nanowires create a network topology like a biological brain
network. Each metal nanowire is covered in a thin insulating layer, similar
to how neurons have an insulating membrane.
Ions flow over the insulating layer and into a nearby nanowire when we
excite nanowires with electrical impulses (much like neurotransmitters
across synapses). As a result, we see electrical signaling in nanowire
networks that resembles synapses.
Memory and learning
Our most recent research investigates the possibility of human-like
intelligence via a nanowire technology. Two characteristics that are
suggestive of high-order cognitive function—learning and memory—are at the
heart of our inquiry.
Our research indicates that synaptic pathways in nanowire networks may be
selectively strengthened (and weakened). In the brain,
"supervised learning"
is comparable to this.
The output of synapses is compared to a desired outcome in this procedure.
When the output of the synapses is close to the intended outcome, they are
either reinforced or pruned, depending on whether they are not.
We built on this finding by demonstrating that by "rewarding" or
"punishing" the network, we could enhance the amount of strengthening. The
brain's "reinforcement learning" serves as the inspiration for this
procedure.
We also utilized a variation of the "n-back task," a test that evaluates
working memory in people. A succession of stimuli are presented, and each
new entry is compared to one that happened n steps before.
Prior signals were "remembered" by the network for at least seven steps.
Strangely, it's commonly believed that people can hold seven things in
working memory at once.
The network's memory performance significantly improved when we employed
reinforcement learning.
We discovered in our nanowire networks that how synapses were previously
stimulated affects how synaptic pathways are formed. The brain's synapses
exhibit this as well; neuroscientists refer to this as "metaplasticity" in
this context.
Artificial intelligence Replicating human intellect is probably still a
long way off.
Our study of neuromorphic nanowire networks, however, demonstrates that it
is feasible to include qualities necessary for intelligence—such as learning
and memory—in non-biological, physical technology.
Nanowire networks are distinct from the AI-related artificial neural
networks. Nevertheless, they could result in so-called "synthetic
intelligence".
One day, maybe, a neuromorphic nanowire network will be able to recall
discussions that are more like human interactions than ChatGPT.
Alon Loeffler, PhD researcher,
University of Sydney
and
Zdenka Kuncic, Professor of Physics,
University of Sydney
