Samsung wants to 'copy and paste' a brain onto memory chips

[nanoguy]

In brief: Samsung is currently researching ways to download the brain's neuronal map onto an electronic device. That kind of technology is likely several years away from coming to fruition, but it looks like Samsung is just as interested in making memory chips as it is making silicon that mimics brain function.

Samsung's latest ambition is to take a new approach in developing chips that mimic the human brain function for self-learning tasks. Specifically, the company wants to do so by copying the real thing onto special silicon.

To explore this new possibility, the Samsung Advanced Institute of Technology (SAIT) partnered with researchers from Harvard University. The preliminary findings are published in a paper in Nature Electronics titled "Neuromorphic electronics based on copying and pasting the brain."

The idea proposed in the new research is that it should be possible to copy the brain's neural connection map using a specially-developed nanoelectrode array developed by Professor HongKun Park and Professor Donhee Ham. The destination for pasting this map is a high-density 3D network of solid-state memory cells, which could be either standard NAND used in SSDs or more exotic memory like resistive RAM.

When pasting the neural map recordings on something like resistive RAM, each cell would be programmed so that its conductance represents the strength of a certain neural connection. However, this is no small task, as the human brain has an estimated 100 billion neurons with many more synaptic connections between them.

This makes the process rather slow and complex, but Samsung says memory technology has evolved to the point where it can effectively help in reverse-engineering the human brain on memory chips. Ultimately, this would allow researchers to more easily develop neuromorphic chips with the same characteristics as the brain, such as low-power operation, quick adaptation to a given environment, and even cognition.

Permalink to story.

https://www.techspot.com/news/91435-samsung-wants-copy-paste-brain-onto-memory-chips.html

 

[VitalyT]

Samsung is currently researching ways to download the brain's neuronal map onto an electronic device. That kind of technology is likely several years away from coming to fruition

I don't get it. Neural networks, built on tensors have been part of nVidia GPU-s for a number of years now. I don't understand what Samsung is trying different here. Reconstructing tensors as physical elements of the chip? Why on earth? Physical resemblance of the modern neural network constructs won't be any better than what we have today, which is highly optimized for best performance, while performance of our brain elements is limited by what nature gave us. This is why it has always been part of many science predictions that humans someday will have chips integrated into the brain, not brain integrated into chips. Samsung seems to be lost here completely.

 

[wiyosaya]

I don't get it. Neural networks, built on tensors have been part of nVidia GPU-s for a number of years now. I don't understand what Samsung is trying different here. Reconstructing tensors as physical elements of the chip? Why on earth? Physical resemblance of the modern neural network constructs won't be any better than what we have today, which is highly optimized for best performance, while performance of our brain elements is limited by what nature gave us. This is why it has always been part of many science predictions that humans someday will have chips integrated into the brain, not brain integrated into chips. Samsung seems to be lost here completely.

Maybe Samsung is aiming to download the contents of a human brain to a chip.

 

[Revolution 11]

Ghost in the Shell, Gangnam Style.

 

[Avro Arrow]

I can see it now...

"WHO WANTS TO BE THE FIRST REAL-WORLD LAWNMOWER MAN?"

 

[Puiu]

I don't get it. Neural networks, built on tensors have been part of nVidia GPU-s for a number of years now. I don't understand what Samsung is trying different here. Reconstructing tensors as physical elements of the chip? Why on earth? Physical resemblance of the modern neural network constructs won't be any better than what we have today, which is highly optimized for best performance, while performance of our brain elements is limited by what nature gave us. This is why it has always been part of many science predictions that humans someday will have chips integrated into the brain, not brain integrated into chips. Samsung seems to be lost here completely.

They want to achieve self-learning that mimics human brains. If they manage to do that then you could theoretically be able to create AI with knowledge from humans and with the ability to improve by themselves.

 

[sagbobbit]

"100 billion neurons"
More like ~85 bil

 

[Plutoisaplanet]

"100 billion neurons"
More like ~85 bil

After smoking whatever Samsung is, I'm sure I could easily lose 15 billion neurons

 

[Guberian]

Each day we get closer to the matrix.

 

[guessedaccount]

you could theoretically be able to create AI with knowledge from humans and with the ability to improve by themselves.

I see no evidence humans are improving by themselves

 

[Hardware Geek]

I can see it now...

"WHO WANTS TO BE THE FIRST REAL-WORLD LAWNMOWER MAN?"

That movie freaked me out as a kid. Now I want to find it and watch it again. Heck, with how many remakes of old movies that come out, I'd definitely watch a new lawnmower man with today's CGI capabilities.

 

[Hardware Geek]

I see no evidence humans are improving by themselves

Agreed. Technology may have improved, but society seems to be tearing itself apart even faster than ever.

 

[Scrye74]

Does this mean my PC will soon be offended by everything?

 

[Avro Arrow]

Agreed. Technology may have improved, but society seems to be tearing itself apart even faster than ever.

This is what happens when you give great power to people who aren't properly taught how to wield it. It's like giving a car to someone without a licence. Nobody stopped to remember how stupid most people still are and how bad some of the so-called "first-world" educations systems are.

 

[Hardware Geek]

This is what happens when you give great power to people who aren't properly taught how to wield it. It's like giving a car to someone without a licence. Nobody stopped to remember how stupid most people still are and how bad some of the so-called "first-world" educations systems are.

Well put. The education system in some first world countries (what an arrogant term when you think about it) is absolutely broken and is only getting worse with time.

 

[cliffordcooley]

Have I missed the part where a single memory was downloaded? How do they think they are going to skip this and go straight to the entire brain?

 

[bobc4012]

Does this mean my PC will soon be offended by everything?

Your post made me think of "Person of Interest". If you never saw the show, you may be able to find it on a streaming site.

 

[Puiu]

Have I missed the part where a single memory was downloaded? How do they think they are going to skip this and go straight to the entire brain?

I don't think they are targeting memories, at least not now.

 

[cliffordcooley]

Then what is it they hope to achieve by copying random bits?

 

[Puiu]

Then what is it they hope to achieve by copying random bits?

Mimic brain functions.

From the article:
"This would allow researchers to more easily develop neuromorphic chips with the same characteristics as the brain, such as low-power operation, quick adaptation to a given environment, and even cognition."

But in general I think they want to mimic how the brain sends and receives information, how it organises information, etc. It could help in the future create chips that might help those that have brain injuries, nerve/spine damage, etc.

 

[cliffordcooley]

I've got to be under thinking it. Seems to be way over my head. I will wish them luck and call it a day.

 

[Avro Arrow]

Well put. The education system in some first world countries (what an arrogant term when you think about it) is absolutely broken and is only getting worse with time.

Agreed on both counts. That's why I put it in quotes. It's really hard to nail down because sometimes I say NATO Countries or the Western World but they leave out Australia which has the same problems. I'm willing to bet that the education systems of the Nordic countries and New Zealand are probably top-notch. Of course, this is why the US government and media do their best to ignore the fact that those countries exist.

 

[enemys]

I don't get it. Neural networks, built on tensors have been part of nVidia GPU-s for a number of years now. I don't understand what Samsung is trying different here. Reconstructing tensors as physical elements of the chip?

Regular artificial neural networks are vastly different from actual biological neurons. They are very useful, but don't actually model how neurons in a brain work. Samsung doesn't want to "reconstruct tensors", more like they want to model biological neurons in hardware.

Physical resemblance of the modern neural network constructs won't be any better than what we have today, which is highly optimized for best performance

But it's not about reconstructing neural networks in hardware or accelerating them, there's tons of hardware startups and big corps doing that. It's about researching asynchronous neuromorphic computing. It's a very different approach to computing and it doesn't translate well to what we use today, even to artificial neural networks. It can solve similar problems and it is a kind of machine learning, but functions very differently.

while performance of our brain elements is limited by what nature gave us. This is why it has always been part of many science predictions that humans someday will have chips integrated into the brain, not brain integrated into chips.

First of all, human brain's performance as a neural network is actually enormous. Second of all, it's not about performance, but about adapting the asynchronous neuron-like computation model.

Samsung seems to be lost here completely.

I would say that you are lost, since you don't understand the difference between neuromorphic models (https://en.wikipedia.org/wiki/Neuromorphic_engineering), e.g. spiking neural networks (https://en.wikipedia.org/wiki/Spiking_neural_network) and regular artificial neural networks (https://en.wikipedia.org/wiki/Artificial_neural_network - see criticism for differences vs biological networks). Classic NNs are doing great in multitudes of scenarios and are developing fast, but this is about another approach to computing, more closely resembling biological neurons. This has become an area of interest for hardware and software manufacturers lately, e.g. Intel has been researching it for years now: https://www.anandtech.com/show/16960/intel-loihi-2-intel-4nm-4.

 

[cliffordcooley]

I would say that you are lost, since you don't understand the difference between neuromorphic models (https://en.wikipedia.org/wiki/Neuromorphic_engineering), e.g. spiking neural networks (https://en.wikipedia.org/wiki/Spiking_neural_network) and regular artificial neural networks (https://en.wikipedia.org/wiki/Artificial_neural_network - see criticism for differences vs biological networks).

I don't know about VitalyT. That would be why I'm lost on this topic.