The tech company is proposing a way to copy a brain’s neuron wiring map into a 3D neuromorphic chip. The approach, detailed in a new paper, relies on a nanoelectron array developed by a group of researchers that enters a big volume of neurons to record where the connections are made and the strength of those connections. Researchers believe that with this approach, they will one day be able to download people’s brains onto 3D chips.
Neuro-morphism is a term we may want to become familiar with. It refers to something that takes the form of the brain. If you want to build something that follows the shape of the brain, you also have to understand the mechanisms of the brain — from why it remembers information to how many neurons are activated before a decision is made. A neuromorphic device could work in a way that’s mechanically analogous to our understanding of some part of the brain.
From Google to Microsoft, many organizations are working to develop neuromorphic chips. Researchers at MIT even discovered (last year) how to put thousands of artificial brain synapses on a very small chip. Even billionaire Elon Musk recently took some big steps with his Neuralink tech company as it builds a device to embed on a person’s brain.
Working with engineers from Harvard University, Samsung presented an approach to create a memory chip that approaches the computing features of the brain that have so far been outside the reach of current technology. This includes autonomy, cognition, low power, facile learning, and adaptation to the environment, for example, all developed around a brain-like design.
“The vision we present is highly ambitious,” Donhee Ham, Fellow of Samsung Advanced Institute of Technology (SAIT) and Professor of Harvard University, said in a media statement from Samsung. “But working toward such a heroic goal will push the boundaries of machine intelligence, neuroscience, and semiconductor technology.
The human brain contains roughly 86 billion neurons, and the way they are connected is even more complex. These connections are largely behind the functions of the brain and they’re what make it so special as an organ. For neuromorphic engineering, the goal has always been, at least since the field officially started in the 1980s, to mimic the structure and function of the neuronal network on a silicon chip.
Nevertheless, this has proven more difficult than expected, as there’s not that much knowledge on how neurons are linked together to create the brain’s higher functions. That’s why the original target of neuromorphic engineering was been recently changed to design a chip “inspired” by the brain instead of trying to mimic it so rigorously.
However, the researchers at Samsung are now suggesting a way to back to the original goal of neuromorphics. The nanoelectrode the developed would enter a big number of neurons and register their electric signals with a high level of sensitivity. The recordings would then inform where neurons connect with each other and the strength of those connections.
The neuronal wiring map could be copied based on those recordings and then pasted to memories, either non-volatile ones, such as those commercially available in solid-state drives (SSD), or to recently developed ones, such as resistive random-access ones (RRAM). Each memory would represent the strength of the neural connection in the map.
In their paper, the researchers also suggest a way to paste the neuronal map to a memory network, using specially-engineered non-volatile memories. These can learn and express the neuronal map when driven by intercellularly recorded signals. There’s a challenge, however, as the human brain has 100 billion neurons, and a thousand times more synaptic connections.
This means the chip will require about 100 trillion or so memories, a difficult challenge for Samsung. While the researchers are optimistic that they could use a 3D integration of memories to address this issue, it will probably take quite a bit of time for Samsung or any other company working on neurophormism to further implement the technology.
The study behind the technology was published in the journal Nature.