Samsung engineers, together with scientists from Harvard University, proposed a new idea that will bring the technological world one step closer to creating neuromorphic chips that mimic the work of the brain. The details were published in Nature Electronics, titled Neuromorphic Electronics Based on Brain Copy and Paste.
The essence of the concept put forward by the authors is really most accurately conveyed by the words "copy" and "paste". The article proposes a method for copying the neural connections of the brain using an array of nanoelectrodes, developed by Harvard University professor and SAIT (Samsung Advanced Institute of Technology) Donhee Ham, as well as by Harvard University professor Hongkun Park. The card created in this way is then inserted into the solid-state memory 3D network module - in this area Samsung is one of the world leaders.
Using this copy-and-paste approach, the authors of the project intend to create a memory microcircuit that will allow to achieve the unique capabilities of the brain: low energy consumption, easy learning, adaptation to the environment, as well as autonomy and the ability to learn - all this was previously inaccessible to computing systems.
The brain consists of a large number of neurons, and the circuitry of their interconnections is responsible for its work. So knowing an accurate brain map turns out to be the key to reverse engineering the entire brain. The branch of neuromorphic engineering originated in the eighties of the last century, at that time its goal was to imitate the structure of the brain and the function of neural networks on a silicon chip. The task seemed extremely difficult, since until now it was almost impossible to make an accurate map of the neural connections responsible for higher brain functions. Therefore, the task of neuromorphic engineering was simplified to the development of a chip "inspired" by the brain, but not replicating it in detail.
The authors of the new project proposed a way to return to the original goal of reverse engineering the brain. An array of nanoelectrodes connects to a large number of neurons and records their electrical signals with high sensitivity. Based on these records, a map of neural connections is created, showing where neurons connect to each other, as well as how strong such a connection is. As a result, a neural wiring map is extracted or "copied" from these records.
The neural map copied in this way is then "inserted" into the nonvolatile memory network - this can be traditional commercial flash memory, which is used in everyday life in solid state drives (SSD), or new generation RRAM memory, where the conductivity of each element corresponds to the strength of the neural connection in previously copied map.
The article also proposes a way to quickly "insert" a map of neural connections into an electronic memory network. A network of specially designed non-volatile memory elements ultimately reflects a map of neural connections, and it is activated by means of fixed intercellular signals. In other words, the neural connections are directly unloaded into the memory chip.
The human brain contains about 100 billion neurons and about a thousand times as many synaptic connections, so in the end the neuromorphic chip will consist of about 100 trillion elements. The integration of such a large number of elements is made possible by Samsung's own 3D technology.
Leveraging its best practices in electronic components manufacturing, Samsung plans to continue working in neuromorphic engineering to consolidate its leadership in semiconductor solutions focused on artificial intelligence.
“The vision we have presented is very ambitious. But working on such a heroic challenge will push the boundaries of machine intelligence, neuroscience and semiconductor technology , ”said Professor Ham.
0 Comments:
Post a Comment
Your comment and facebook share will be appreciated