Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

A mechanical neural network learns to respond to its environment

hackster.io

Post reply on HN
Print view View original

11–19 of 19 posts

Re: A mechanical neural network learns to respond to its environment

#11

Post by djokkataja »

bufferoverflow wrote: post #7
iNic wrote: post #2

I don't understand the advantage of this over just having a digital NN and then pushing the output to some electric motor?

I think you're missing the point. I think it's to demonstrate that fully mechanical systems can learn and react like a real brain plus muscles and sensors. The possible implication is, if consciousness can arise in a digital neural net, it can also arise in a mechanical system, you just need enough of levers and gears connected in a certain way.

> I think it's to demonstrate that fully mechanical systems can learn and react like a real brain plus muscles and sensors.

That wasn't the point of the research as expressed in the article (note that the final image has a label for "Control electronics" underneath the mechanical system):

> Thus, this work lays the foundation for artificial-intelligent (AI) materials that can learn behaviors and properties.

With applications such as:

> use in aircraft wings to morph the shape in response to wind patterns to boost efficiency, adding reactive rigidity to buildings to better withstand earthquakes and other disasters, shockwave-deflecting reactive armor, or even the creation of surfaces able to perform acoustic imaging.

Re: A mechanical neural network learns to respond to its environment

#12

Post by abecedarius »

andrewflnr wrote: post #10

Given the title "Mechanical Neural Network", the presence of a conventional digital computer is really disappointing. I went in to this article hoping for steampunk AI.

Here's an idea: ReLU basically = a seesaw + ropes that attach at a movable distance from the fulcrum. Adjusting a 'weight' = sliding the attachment. Nonlinearity = rope can pull but not push.

I'm afraid getting the details to work in practice would probably be horrible.

Not steampunk but sort of biopunk: years ago someone programmed a DNA strand displacement mechanism to implement a particular fixed neural network, as a proof of concept.

Re: A mechanical neural network learns to respond to its environment

#14

Post by betwixthewires »

djokkataja wrote: post #11
bufferoverflow wrote: post #7

Earlier quoted context omitted.

I think you're missing the point. I think it's to demonstrate that fully mechanical systems can learn and react like a real brain plus muscles and sensors. The possible implication is, if consciousness can arise in a digital neural net, it can also arise in a mechanical system, you just need enough of levers and gears connected in a certain way.

> I think it's to demonstrate that fully mechanical systems can learn and react like a real brain plus muscles and sensors. That wasn't the point of the research as expressed in the article (note that the final image has a label for "Control electronics" underneath the mechanical system): > Thus, this work lays the foundation for artificial-intelligent (AI) materials that can learn behaviors and properties. With appl…

Well of course they have useful applications in mind. But it does demonstrate something interesting: computation occurs naturally, on the "bare metal" of the universe, and systems can be built that perform computation that do not require an abstraction or purpose built computing device.

Re: A mechanical neural network learns to respond to its environment

#15

Post by betwixthewires »

akihitosan wrote: post #8

I think to call a big machine a material is very wrong.

Why? Isn't the distinction somewhat arbitrary? A lattice or scaffold is indistinguishable from a crysralling structure only due to it's scale, but when a large, rigid ststructure is built it functions on a larger scale as a sort of material.

Re: A mechanical neural network learns to respond to its environment

#16

Post by bitwize »

Snake: Reactive armor, huh?

Otacon: That's right, Snake. The material you're wearing is a mechanical neural network -- meaning there's no computer chips or other SPOF (single point of failure). The material itself is smart, the fabric is doing the thinking as a whole, responding to the environment to protect you from incoming fire.

Snake: I don't like the feeling of wearing something intelligent.

Otacon: Just think of it as like an alien symbiote costume, like the heroes from those old tokusatsu shows wore.

Snake: You're not helping.

Re: A mechanical neural network learns to respond to its environment

#17

Post by mach1ne »

juunpp wrote: post #6
mach1ne wrote: post #5

>Learns to Respond to Its Environment So we're back at this kind of antropomorphization.

*Anthropos, and as if learning was unique to humans? What is learning anyway if not the adaptation of the brain to its inputs?

Saying the system learns to 'respond to its environment' is an overly generous description which muddles up what is actually happening.

Re: A mechanical neural network learns to respond to its environment

#18

Post by xg15 »

abecedarius wrote: post #12
andrewflnr wrote: post #10

Given the title "Mechanical Neural Network", the presence of a conventional digital computer is really disappointing. I went in to this article hoping for steampunk AI.

Here's an idea: ReLU basically = a seesaw + ropes that attach at a movable distance from the fulcrum. Adjusting a 'weight' = sliding the attachment. Nonlinearity = rope can pull but not push. I'm afraid getting the details to work in practice would probably be horrible. Not steampunk but sort of biopunk: years ago someone programmed a DNA strand displacement mechanism to implement a particular fixed neural network, a…

That's a good start, but I'd say it's about as far as the people in the article got - i.e. if you already know the weights, you can do inference fully in mechanics.

That could be quite useful for pretrained stuff, no question. But I imagine, to go full steampunk, you'd also have to find a way to train the network fully mechanically.

I.e. implement something like gradient descent without using a digital computer...

Re: A mechanical neural network learns to respond to its environment

#19

Post by abecedarius »

xg15 wrote: post #18
abecedarius wrote: post #12

Earlier quoted context omitted.

Here's an idea: ReLU basically = a seesaw + ropes that attach at a movable distance from the fulcrum. Adjusting a 'weight' = sliding the attachment. Nonlinearity = rope can pull but not push. I'm afraid getting the details to work in practice would probably be horrible. Not steampunk but sort of biopunk: years ago someone programmed a DNA strand displacement mechanism to implement a particular fixed neural network, a…

That's a good start, but I'd say it's about as far as the people in the article got - i.e. if you already know the weights, you can do inference fully in mechanics. That could be quite useful for pretrained stuff, no question. But I imagine, to go full steampunk, you'd also have to find a way to train the network fully mechanically. I.e. implement something like gradient descent without using a digital computer...

Yeah, for backprop you need an error signal (probably not a rope now because it's signed), and to multiply it by the activation (which suggests another seesaw for the multiply), and to bump the connection horizontally proportional to the product. Then fan out the product into the previous layer's error.

I have no experience at mechanical design -- I got to thinking about this out of irritation at a pop-science explanation of neural networks in the New York Times magazine which explained nothing. Could you help people to a real understanding (a basic one) by letting them play with a tiny network in tangible form, where you can see the activations and feel the forces?

I guess making it work with all the details and practicalities would probably end up with the basic parts too complicated, and too many of them for this purpose of learning, but it'd be cool to try.

Post reply on HN