How to make brain-like chips? Samsung & Harvard: Directly copy and paste neurons | Nature sub-publication

Fengse from Aofei Temple
Quantum Bit Report | Public Account QbitAI

How to make brain-like chips ?

Recently, Samsung and Harvard University proposed a "simple and crude" method:

copy and paste .

It is also reported that this research is beyond the reach of current technology, and therefore it was recently published in a Nature journal.

Specifically, it is done by "copying and pasting" the brain's neuronal connection map onto the high-density 3D storage network of solid-state memory.

In this way, it is possible to realize chips that are low-power, easy to learn, adaptable to the environment, and even have characteristics such as autonomy and cognitive capabilities.

So why did they do such a study?

Samsung said that although neuromorphic electronics was proposed as early as the 1980s, it is too difficult to imitate the human brain .

So what everyone has been doing has been simplified into technologies that are "inspired" by the brain rather than strict imitations.

Now, they have returned to their original vision of truly mimicking the structure and function of the human brain's neuronal networks on silicon integrated circuits.

How to "copy and paste"?

To achieve this goal, the researchers used nanoelectrode arrays and memory chips to "copy and paste" neuronal connections.

The nanoelectrode array here adopts the previous research results of the team at Harvard University: CMOS nanoelectrode array (CNEA) .

The CMOS nanoelectrode array integrates 4096 electronic channels into a CMOS chip with 4096 vertical nanoelectrodes to connect the parallel recording of intracellular recording (an important research direction in neuroscience) and map the functional synaptic connections of neurons.

The left picture below shows rat neurons simulated by CMOS nanoelectrode array under electron microscope;

On the right is the synaptic connection map extracted from the CNEA using a computer-assisted analysis program.

After “copying” or extracting the neuron connection diagram, the next step is to “paste” it into the storage network.

This can be achieved by programming a synthetic memory network so that each memory can reflect the strength of each neuron connection through conductance .

Alternatively, you can also download the connection diagram directly to the memory chip:

The recorded signals are used to directly drive an N×N memory crossbar array, performing physical imprinting of the synaptic connection graph .

The storage networks that can be carried in the "paste" step include flash memory, magnetic random access memory (MRAM) , phase change random access memory (PRAM) and resistive random access memory (RRAM) .

The above is a brief introduction to the team’s “copy and paste” method. If you are interested, you can check out the paper for further information.

Providing a new possibility and direction for the study of neuronal connections

Although this method is still just a theory, there are still many challenges in actually implementing it.

But this research is a pioneering effort that provides a new possibility and direction for the study of neuronal connections and has great application prospects.

Samsung said it is also leveraging its experience in chip manufacturing, hoping to advance the development of machine intelligence, neuroscience and semiconductor technology through research into neuromorphic engineering.

At the same time, they also made no secret of their "ambitious" pursuit of their leading position in the next generation of artificial intelligence semiconductors .

Oh yes, back to the research itself, since there are about 100 billion neurons in the human brain, and about a thousand times more synaptic connections, such a neuromorphic chip will eventually require about 100 trillion memories.

However, through 3D integration of memory , it will be possible to integrate such a large-scale memory on a single chip.

This 3D integration technology, which has opened up a new era for the memory industry, was also developed by Samsung .

Paper address:
https://www.nature.com/articles/s41928-021-00646-1

Reference link:
https://news.samsung.com/global/samsung-electronics-puts-forward-a-vision-to-copy-and-paste-the-brain-on-neuromorphic-chips?utm_source=rss&utm_medium=direct