Mechanical neural network can learn and change its physical properties | Hacker News Reader

Mechanical neural network can learn and change its physical properties

theconversation.com

73 pointsBrajeshwar14 comments

I think we're stretching the definition of a 'mechanical neural network' when it's driven by an electrical solenoid that is is turn driven from an electronic computer. This article is neat, but the title is kind of clickbait miss-the-point-but-get-the-click journalism.

I can't seem to access the paper itself to see what is really gong on, and if the solenoids could be theoretically replaced with some fancy material, but I think the correct term of the current prototype is "electro-mechanical computer":

https://en.wikipedia.org/wiki/Mechanical_computer#Electro-me...

But this thing, the mechanical part, doesnt do any computation. It just a matrix of sensors and actuators plugged into an off-the-shelf electronic computer.

Right, reading the title, I was assuming the material or machine itself trained nodes based on the stresses or shears it felt itself, but it being computer controlled doesn't really seem to mean it is a "mechanical neural network" in the sense that it "learns." I mean, you could say (from my field) a deformable mirror is a "mechanical neural network" that existed decades before NN became a hot topic.

This concept has an interesting similarity to plant growth, in particular stem development. Strong stem growth in early plant development relies in part on wind exposure, and without any wind, plants will collapse later in development under their own weight (a common problem in greenhouse and indoor cultivation).

https://awesci.com/the-role-of-wind-in-a-trees-life/

However, trees can't reshape themselves. Once a woody structure is formed, it's set. This allows for highly durable structures, and that might be a problem with this technology - i.e. the wear and tear that comes with continually reforming the material. This is solved in animals (i.e. muscle tissue) by continually regenerating the material in an active turnover process, which might be required here as well.

> Strong stem growth in early plant development relies in part on wind exposure, and without any wind, plants will collapse later in development under their own weight

That's interesting - I wonder if there's been any studies to find out if vegetation behind wind turbines grow abnormally. It would seem like taking terawatts of power from the wind across the planet could take some toll (though still probably a huge net improvement over hydrocarbons).

Anekdotal, but I know someone who has an orchard just beneath a win turbine. Trees grow just fine, and in general ground windspeed does not seem affected. The most significant downside seems to be that staff dislikes the constant whoosh-whoosh sounds

same problems had happened in green house, a electric fan is enough.

Any videos of this in action? I couldn't see any linked in the article.

Reminds me of this previous paper.

Deep physical neural networks trained with backpropagation https://www.nature.com/articles/s41586-021-04223-6

I imagine the T-1000 would have been made of something like this, were it real.

This looks more like a 2d lattice of flextures connected by little linear motors. So more like a so called smart material emulation of a planar muscle with more than one axis of motion.

Since I doubt nature does care about distinction between chemical, mechanical and electrical this is a good step towards understanding nature as far as I am concerned.

Imagine a bridge resting on a neural network… one morning it decides to scratch its own back…

Reply on news.ycombinator.com