Samsung proposes a 'copy and paste' approach to transferring the brain onto neuromorphic chips
Samsung has shared a new idea that brings the world one step closer to neuromorphic chips that could better mimic the brain. The idea, developed by leading engineers and scientists from Samsung and Harvard University, was published in Nature Electronics as an article titled "Neuromorphic electronics based on copying and pasting the brain."
The authors of the article are Donhee Ham, a researcher at the Samsung Advanced Institute of Technology (SAIT) and a Harvard University professor; Hongkun Park, a Harvard University professor; Sungwoo Hwang, president and CEO of Samsung SDS and former head of SAIT; and Kinam Kim, vice chairman and CEO of Samsung Electronics.
"Reverse engineering the brain by mimicking the structure and functions of neural networks on a silicon integrated circuit was the original goal of neuromorphic engineering, but it remains a distant prospect. Thus, the focus of neuromorphic engineering has shifted from strict brain emulation to projects inspired by the brain's qualitative characteristics, including event-driven signals and in-memory information processing. Here we examine modern approaches to neuromorphic engineering and propose a vision that returns neuromorphic electronics to its original goal of reverse engineering the brain," the scientific paper says.
The core of the idea is best captured in two words: "copy" and "paste." The article proposes copying the brain's neural connectivity map using a nanoelectrode array developed by Ham and Park, and pasting that map into a three-dimensional high-density solid-state memory network, a technology in which the company is the global leader.
Using this copy-and-paste approach, the authors suggest it may be possible to create a memory chip that approaches the brain's unique computing characteristics: low power consumption, easy learning, adaptation to the environment, as well as autonomy and cognition. These characteristics have always been out of reach for modern technology.
The brain consists of a large number of neurons, and their wiring layout is responsible for the organ's functions. Thus, knowing the map is the key to reverse engineering the brain.
The original goal of neuromorphic engineering, which emerged in the 1980s, was to emulate such neural network structure and function on a silicon chip. This proved difficult, since little is still known about how a large number of neurons are connected to produce the brain's higher functions. Thus, the goal of neuromorphic engineering was simplified to creating a chip "inspired" by the brain rather than precisely imitating it.
The authors propose a way to return to the original neuromorphic goal of reverse engineering the brain. The nanoelectrode array can efficiently penetrate a large number of neurons, so it can record their electrical signals with high sensitivity. These massively parallel intracellular recordings inform the neural wiring map, indicating where neurons connect to each other and how strong those connections are. Consequently, the neural wiring map can be extracted or "copied" from these control recordings.
The copied neural map can then be "pasted" into a network of non-volatile memory, such as commercial flash memory used in everyday life in solid-state drives (SSDs), or "new" memory such as resistive random-access memory (RRAM), by programming each memory cell so that its conductance represents the strength of each neural connection in the copied map.
The article takes a step further and proposes a strategy for rapidly pasting the neural wiring map into a memory network. A network of specially designed non-volatile memory devices can learn and display the map of neural connections when it is directly driven by the intracellular recorded signals. This is a scheme that directly loads the brain's neural connection map onto a memory chip.
Since the human brain has about 100 billion neurons and roughly 1,000 times more synaptic connections, a final neuromorphic chip would require about 100 trillion memories. Integrating such a huge amount of memory on a single chip has become possible thanks to three-dimensional memory integration, a technology led by Samsung that has opened a new era in the industry.
Drawing on its advanced chipmaking expertise, the company plans to continue research in neuromorphic engineering to strengthen its leadership in next-generation artificial intelligence semiconductors.
“The vision we present is very ambitious. But work toward such a heroic goal will push the boundaries of machine intelligence, neuroscience and semiconductor technology”
This is clearly a challenging task for any company, since there is no code needed to run a virtual brain. Samsung may have opened the door to human-like artificial intelligence, but it may be a very long time before anyone reaches that goal.
