Picture it in your mind - Neural chip with new alloy tested

[williamcll]

Scientists have made striding improvements to processors that operate in a manner similar to neurons in a brain. For this experiment, the neural processor was made with a new alloy and could remember images like that of a person as a demonstration. They hopes to apply this chip in mobile situations in the future.

Quote

In a paper published in Nature Nanotechnology, the researchers explain how their brain-inspired chip was able to remember and recreate a gray-scale image of Captain America’s shield and reliably alter an image of MIT’s Killian Court by sharpening and blurring it. Those tests may seem minor, but the team believes the chip design could advance the development of small, portable AI devices and carry out complex computational tasks that today only supercomputers are capable of.

“So far, artificial synapse networks exist as software. We’re trying to build real neural network hardware for portable artificial intelligence systems,” says Jeehwan Kim, associate professor of mechanical engineering at MIT. “Imagine connecting a neuromorphic device to a camera on your car, and having it recognize lights and objects and make a decision immediately, without having to connect to the internet.”

For your information, the processor are made of memristors, which operates in the manner described below:

Quote

 In a neuromorphic device, a memristor would serve as the transistor in a circuit, though its workings would more closely resemble a brain synapse—the junction between two neurons. The synapse receives signals from one neuron, in the form of ions, and sends a corresponding signal to the next neuron. A transistor in a conventional circuit transmits information by switching between one of only two values, 0 and 1, and doing so only when the signal it receives, in the form of an electric current, is of a particular strength. In contrast, a memristor would work along a gradient, much like a synapse in the brain. The signal it produces would vary depending on the strength of the signal that it receives. This would enable a single memristor to have many values, and therefore carry out a far wider range of operations than binary transistors. 

 

Ultimately, scientists envision that memristors would require far less chip real estate than conventional transistors, enabling powerful, portable computing devices that do not rely on supercomputers, or even connections to the Internet.

Source: https://www.engadget.com/mit-brain-on-a-chip-silver-copper-memristors-183405642.html

https://www.nature.com/articles/s41565-020-0694-5

https://phys.org/news/2020-06-tens-thousands-artificial-brain-synapses.html

Thoughts: Impressive, but I don't think a processor like this would replace a CPU anytime soon, at most an accelerator card or part of a phone SoC. I also wonder if this could be used to replace damaged brains since the memristors operate similar to a neuron.

[Nicnac]

i still don't get how they made it remember a picture with some sort of sharpening function post-processing?

 

I do get that basically using an action potential in neuromorphic circuits can be a very interesting and maybe useful application in some instances.

I will read the paper in detail when i have the time for it. Also curious if thats just mimicking a voltage gate or if the current is really truly added across the board if that makes sense...

[Drama Lama]

I don’t understand what the benefits of this instead of traditional picture analyzing software is

[BobVonBob]

6 minutes ago, Drama Lama said:

I don’t understand what the benefits of this instead of traditional picture analyzing software is

From what I understand, this is a hardware implementation of a neural network.

[Drama Lama]

17 minutes ago, BobVonBob said:

From what I understand, this is a hardware implementation of a neural network.

So like an fpga for neural networks?

[BobVonBob]

9 minutes ago, Drama Lama said:

So like an fpga for neural networks?

Seems like it, although the articles aren't all that in depth and I don't have access to the paper other than the abstract.

[Nicnac]

No it’s not. While NNs use synaptic connections in a way the way that these synaptic clefts in a way communicate is different here. There is no 1 or 0 but the current or voltage seems to be transmitted as a whole so it can add up or be reduced at these connections. At least that’s what I gathered from it. 

[ARikozuM]

So, I have to raise my iPhone from infancy to maturity through bedtime stories and awkward "bird and bee" conversations? 

[sub68]

3 hours ago, Drama Lama said:

I don’t understand what the benefits of this instead of traditional picture analyzing software is

Probaly easier to doit

[Void Master]

This can prove really useful....

[Sauron]

4 hours ago, williamcll said:

“So far, artificial synapse networks exist as software. We’re trying to build real neural network hardware for portable artificial intelligence systems,” says Jeehwan Kim, associate professor of mechanical engineering at MIT. “Imagine connecting a neuromorphic device to a camera on your car, and having it recognize lights and objects and make a decision immediately, without having to connect to the internet.”

So it's just an AI oriented ASIC...

 

also wtf, you don't need an internet connection to use a "normal" software neural network maybe don't ask this to the mechanical engineering professor?

[Sauron]

45 minutes ago, Nicnac said:

No it’s not. While NNs use synaptic connections in a way the way that these synaptic clefts in a way communicate is different here. There is no 1 or 0 but the current or voltage seems to be transmitted as a whole so it can add up or be reduced at these connections. At least that’s what I gathered from it. 

Yes, but ultimately it doesn't make a lot of difference - you can do the same by using multiple "normal" transistors. This probably has some efficiency benefits compared to that approach but that's about it. It's still an ASIC by definition (fpga might not quite fit though since it's not really a "gate array" in the traditional sense), how the hardware is implemented ultimately doesn't matter.

32 minutes ago, sub68 said:

Probaly easier to doit

Not easier, just likely faster provided they can refine the process enough to compete with more conventional systems.

[sub68]

3 minutes ago, Sauron said:

Not easier, just likely faster provided they can refine the process enough to compete with more conventional systems.

makes sense

[Doobeedoo]

So a memristor as some type of accelerator hm.

[Rohith_Kumar_Sp]

ELI5 plz

[SpaceGhostC2C]

8 hours ago, Drama Lama said:

I don’t understand what the benefits of this instead of traditional picture analyzing software is

Think of the difference between rendering in the CPU or using a GPU.

You can do it either way, but what kind of CPU, and at which manufacturing cost and power consumption level, would match the performance of the GPU?

 

Now factor in that a CPU and a GPU are more similar than either a CPU or GPU and this device.

In principle, it should conceivably open the door for cheaper and more powerful artificial neural networks. And who know, maybe artificial intelligence one day (what we fashionably call "AI" these days is a set of statistics tools - quite useful in some context, but ultimately a glorified hand calculator, and conceptually as close to "intelligence" as one).

 

 

[Drama Lama]

57 minutes ago, SpaceGhostC2C said:

Think of the difference between rendering in the CPU or using a GPU.

You can do it either way, but what kind of CPU, and at which manufacturing cost and power consumption level, would match the performance of the GPU?

 

Now factor in that a CPU and a GPU are more similar than either a CPU or GPU and this device.

In principle, it should conceivably open the door for cheaper and more powerful artificial neural networks. And who know, maybe artificial intelligence one day (what we fashionably call "AI" these days is a set of statistics tools - quite useful in some context, but ultimately a glorified hand calculator, and conceptually as close to "intelligence" as one).

 

 

So it’s not like all those ai chip startups that are popping up in Silicon Valley like mushrooms

not a general neural network chip 

but a chip for one specific neural network 

example : the general neural network chip is like a gpu

and the silicon from this tech news is like this fpga running Doom

Edited June 11, 2020 by Drama Lama

[mr moose]

9 hours ago, Drama Lama said:

I don’t understand what the benefits of this instead of traditional picture analyzing software is

The whole picture thing is to demonstrate what the chip can do and how it works, not what it was designed to do.    It likely isn't designed with any specific end use in mind other than to be able to apply a more brain like processing ability to problems.