(ORDO NEWS) — Computer neural networks are impressive in their results, but they require powerful electronic devices to work: video cards and AI accelerators. Scientists from the USA decided to go the other way and embodied the neural network in a mechanism consisting of springs of variable stiffness.
Mechanical engineers at the University of California at Los Angeles have developed a mechanical analogue of a computer neural network – a three-dimensional lattice structure of interconnected springs with adjustable stiffness.
By adjusting the stiffness of each spring, such a structure can be trained to respond to changing conditions. As scientists report in the journal Science Robotics, in fact, this is a programmable material that can dynamically redistribute loads and dampen emerging stresses.
Neural networks as such work on the logic of connections between neurons with each other. In artificial neural networks, which are available on computers and smartphones, neurons are stacked in several layers.
Initially, they are randomly connected to each other and the neural network as a whole is useless. Therefore, it is trained: for example, if it is a neural network for image recognition, then millions of photographs and pictures are fed into it.
When the neural network “sees” the picture, some neurons of the first layer are activated, which transmit a signal to some neurons of the second layer, and so on, until the last layer produces a single final signal.
Each picture creates its own three-dimensional “pattern” of layer-by-layer activation of neurons, but pictures with similar content, although they activate slightly different neurons, produce the same final signal. For example, the word “dog” for all pictures with dogs.
This is due to the change in connections between neurons during the training of the neural network. Some neurons strengthen the connection with each other, linking into separate groups.
The bottom line is that these groups of neurons are activated together over and over again when there are some characteristic features of the object in the picture.
And if you give the neural network different images of dogs, the same groups of “doglike” neurons will be activated in it, which allows the neural network to respond with the word “dog”. Therefore, training a neural network (including a computer one) comes down to building connections of the required strength between certain neurons.
The Californian engineers’ idea is simple: if you combine mechanical springs with adjustable stiffness into a three-dimensional multilayer structure, then the stiffness of the springs will play the role of the strength of neural connections.
The researchers went through 200 different structures using computer simulations and came to the conclusion that it is most effective to connect the springs together into triangular pyramids. It is this structure that behaves as similar as possible to a primitive computer neural network.
It was difficult for scientists to immediately create a multilayer structure of springs, so in the experiment they limited themselves to a flat lattice of triangles about 60 by 45 centimeters in size, where the size of one spring was about 15 centimeters.
Also, it was not possible to do without electricity at all: the stiffness of the springs in the pilot plant was regulated by the magnetic field of ordinary voice coils, and the strain sensor recorded the force on the spring.
Therefore, the prototype, strictly speaking, should be called an electromechanical neural network, and not a purely mechanical one.
Even such a small and simple structure was enough to train her to respond to changing conditions. The spring system dynamically adjusted to the pressure of different forces coming from different directions in order to reduce it to a load of constant magnitude and direction.
The mechanical neural network of the Californian scientists thus behaved like a “smart” material, the reactions of which can be programmed by training. If you make, for example, an airplane wing out of it, it will be able to adapt to air flows, arching in the right way to prevent strong fluctuations in lift.
And armor from such springy triangles will concentrate maximum strength at the point of impact of the projectile.
True, for this it is necessary to radically reduce the scale of the mechanical neural network and transfer it to the microcosm. The authors of the study propose to use the latest materials with adjustable stiffness for this.
Then, in their opinion, it will be possible to create already three-dimensional lattices from a huge number of springy elements, and the intelligence of such a “real neural network” will increase many times – it will be possible to train much more complex behavior.
Contact us: [email protected]