Samsung and Harvard University have published new research that proposes it is feasible to develop a brain-inspired memory chip. Samsung Electronics has shared a new understanding that makes the world a step nearer to acknowledging neuromorphic chips that can more readily copy the brain.
Imagined by the leading engineers and researchers from Samsung and Harvard University, the insight was published on Sept. 23 as a Perspective paper, named ‘Neuromorphic electronics based on copying and pasting the brain’, by Nature Electronics.
The co-corresponding authors of the paper are Ham Don-hee, a fellow of Samsung Advanced Institute of Technology (SAIT) and professor of Harvard University, professor Park Hong-Kun of Harvard University, Hwang Sung-woo, president and CEO of Samsung SDS and former head of SAIT, and Kim Ki-Nam, vice chairman and CEO of Samsung Electronics.
Samsung thinks it has a superior method to develop brain-like chips: get existing brain structures. The tech firm has proposed a strategy that would “copy and paste” a brain’s neuron wiring map to 3D neuromorphic chips. The methodology would depend on a nanoelectrode array that enters large volumes of neurons to record both where the neurons associate and the strength of those connections. You could copy that information and ‘paste’ it to a 3D network of solid-state memory, regardless of whether it’s off-the-shelf flash storage or cutting-edge memory like resistive RAM.
Every memory unit would have a conductance that reflects the strength of every neuron connection in the map. The result would be a viable re-visitation of “reverse-engineering the brain” like researchers initially wanted, Samsung said.
The move could serve as a ‘shortcut’ to artificial intelligence systems that act like real brains, including the adaptability to learn new concepts and adapt to changing conditions. You may even see completely autonomous machines with true cognition, as indicated by the researchers.
There’s a glaring issue with complexity, nonetheless. As a human brain has approximately 100 billion neurons with multiple times more synaptic links, an ideal neuromorphic chip would require around 100 trillion memory units. That is plainly difficult for any organization, and that does exclude the code expected to make this virtual brain work. Samsung might have made a way for human-like AI, however, it could take a long time before anybody arrives at that objective.
In a point of view paper published in Nature Electronics, the researchers from the banding together associations recommended that the brain’s neuronal connection map could be copied using a nanoelectrode array. They determined that the nanoelectrode array could be used to record the electrical signals created by the enormous number of neurons found in the brain. The recordings could then be used to illuminate the neuronal map by showing where neurons connect and how strong the connections are, the analysts claimed.
When copied, the neuronal map could then be pasted onto a high-density three-dimensional network of solid-state memory, for example, commercial flash memory used in solid-state drives or resistive RAM.
At last, the memory chip would contain traits of the brain, like low power, facile learning, adaption to the environment, autonomy, and cognition, the scientists said.
The paper likewise proposes one potential approach to accelerate pasting the neuronal map is by straightforwardly downloading the map onto a memory chip.
“The vision we present is highly ambitious,” Samsung Advanced Institute of Technology fellow Donhee Ham. “But working toward such a heroic goal will push the boundaries of machine intelligence, neuroscience, and semiconductor technology.”
Looking forward, Samsung plans to proceed with its research into neuromorphic engineering as part of its improvement of AI semiconductors.
During the last quarter, the South Korean tech giant recorded its best-operating profit in almost three years, posting 63.67 trillion won in sales and 12.57 trillion won in operating income.
The substance of the vision set forward by the authors is best summarized by the two words, ‘copy’ and ‘paste.’ The paper recommends an approach to copy the brain’s neuronal connection map using a breakthrough nanoelectrode array created by Ham and Park, and to paste this map onto a high-density three-dimensional network of solid-state memories, the technology for which Samsung has been a world leader.
Through this copy and paste approach, the authors imagine making a memory chip that approximates the remarkable computing traits of the brain – low power, facile learning, adaptation to the environment, and even autonomy and discernment – that have been beyond the reach of current technology.
The brain is made up of a large number of neurons, and their wiring map is responsible for the brain’s functions. In this way, the information on the map is the way to reverse-engineering the brain.
While the original goal of neuromorphic engineering, launched during the 1980s, was to copy such structure and function of the neuronal networks on a silicon chip, it demonstrated troublesome because, even as of recently, little is known about how the huge number of neurons are wired together to make the brain’s higher functions. In this manner, the goal of neuromorphic engineering has been facilitated to designing a chip ‘inspired’ by the brain as opposed to thoroughly emulating it.
This paper proposes an approach to get back to the original neuromorphic objective of the brain reverse engineering. The nanoelectrode array can adequately enter an enormous number of neurons so it can record their electrical signals with high affectability. These hugely parallel intracellular recordings inform the neuronal wiring map, showing where neurons connect and how strong these connections are. Thus, from these telltale recordings, the neuronal wiring map can be removed, or ‘copied’.
The copied neuronal map would then be able to be ‘pasted’ to a network of non-volatile memories –, for example, commercial flash memories that are used in our regular day to day life in solid-state drives (SSD), or ‘new’ memories like resistive random access memories (RRAM) – by programming every memory so its conductance represents the strength of each neuronal connection in the copied map.
The paper makes a step further and recommends a system to quickly paste the neuronal wiring map onto a memory network. A network of specially-engineered non-volatile memories can learn and express the neuronal connection map when straightforwardly driven by intracellularly recorded signals. This is a scheme that straightforwardly downloads the brain’s neuronal connection map onto the memory chip.
Since the human brain has an expected 100 billion or something like that neurons, and a thousand or so times more synaptic connections, an ultimate neuromorphic chip will require 100 trillion or thereabouts memories. Integrating such a huge number of memories on a single chip would be made conceivable by 3D integration of memories, the technology led by Samsung that opened up a new period for the memory industry.
Leveraging its leading involvement with chip manufacturing, Samsung is intending to proceed with its research into neuromorphic engineering, to expand Samsung’s leadership in the field of next-generation AI semiconductors.
“The vision we present is highly ambitious,” Ham said. “But working toward such a heroic goal will push the boundaries of machine intelligence, neuroscience, and semiconductor technology.”
The organization’s semiconductor business contributed 6.93 trillion won in operating income, over half the total figure.
Market conditions worked on in the memory market, which prompted a stronger-than-expected expansion in average sales costs for both DRAM and NAND chips, Samsung said.