待翻譯：AI on a Piece of Sponge?

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Letters imprinted on the same piece of sponge in it’s compressed region (shown by the darker region) In 2020, I was grappling with a piece of sponge. I had come across an unusual effect in a soft elastic material. I published in Soft Matter and its’ older version was available on Arxiv from around a couple of years ago. According to this effect, you can encode memory (of deformation) in the individual rods of the sponge at a millimeter scale. Each orientation of a rod can correspond to a different n-bit, based on how we define it. Thus we can encode a lot of numbers, a lot of information into the piece of sponge. It is not memory foam, because it is reversible. We first put a global stress onto the material to collapse the individual rods. Then at second step we indent them externally. The external indentation stays in place, locally deforming the orientation of rods, thus storing “memory”. If we ease the global stress, our gets erased. If you have a piece of sponge around you, you can play with it. You press it with your finger and it will come back when you remove it. Nothing special, which is why the extra step of applying a global stress first. Now, when we discover something new, we don’t know what the purpose of it is, or would be. Our job as a scientist is to simply report it (hopefully publicly) and let the world do it’s job. If it is worthwhile, take it forward. If it is not, then let it gather dust on a server somewhere. However, when we were writing the paper, my guide constantly asked me, what could be the purpose of such a new form of “memory”‘, and if it is memory at all in the first place. In order to find applications for such a thing, I started looking at computers and how the memory is stored, and has been stored historically. If you zoom in on a Compact or Blu-ray disc, its’ just lines. Lines etched via a laser and it is our ability to associate meaning to those lines that we enable ourselves to encode movies, games and software on them. The meaning we associate with them is digital; binary; 2-bits. However, not all forms of memory have to be binary. Especially music has traditionally been analog, on magnetic tapes and vinyl records, or in fact, IMAX films. Analog means that our ability to write on something or read from it is limited by our own capabilities. That line on the compact disc, could be 6.032455992929… microns long. But our capability to read them can only help us decide if it is smaller, than lets’ say 3 microns (for example, and associate a 0 with it), or longer and associate the other bit with it. If we can distinguish between other sets of lengths, lets’ say 1.5, 3, 4.5 and 6, then we can as well associate 4 numbers to each line, and encode double the memory on the same disc. So, this made me realize, first of all, that mechanical memory is indeed a thing, and it can have different kinds of applications than the electronic memories on our SSDs. The other thing, like I explained previously, is that it is reversible. So, probably slightly more advantageous in the context of non-programmable mechanical memories. Reproducing a figure from the paper that shows buckled memory states. The other thing I found while exploring the possible applications, was finding out about neuromorphic computing. It’s’ the kind of computing in which you combine logic gates with memory instead of keeping them separate as CPU and RAM. Its’ called neuromorphic because it is inspired from neurons which do the same. Each neuron can take signals from different dendrites around it and initiate a signal in its’ output. This output is positive (firing is triggered) if the weights of the dendrites along the positive input signal are large enough or not. The ‘weights‘! These are the same kind of weights that you have in your AI systems, since AI systems were indeed inspired from our brain. So, coming back to my sponge, could this be it? Could the brain be storing the weights in some kind of deformation in its’ own soft fibers? Maybe, maybe not. It’s just a conjecture. However, from what I have learnt during my years of answering such questions, there are two ways of doing science. The first approach is the discovery approach. We try to find out mechanisms of phenomena. This is how it works, this is what controls that. The other way is the engineering approach. I don’t know what the mechanism of how brain works, but because our experiments prove that we have this control ability, let us see what we can build something new out of it. Last time I thought more of neuromorphic computing was in 2020 while working on this project. The next time is today, when I found out about Taalas building architectures which are again, a combination of logic gates and memory. So, how do I propose a sponge could be useful here? Here’s my proposal: Since resistances, capacitances etc are electronic components that change their values depending on the geometry of the resistor or the capacitor, we could design components based on these hyper-elastic materials that can be programmed to different values and reset when needed. And therefore store a form of memory in them that can be read by the electronic circuit as well, and change it’s processing and output depending on the programmed component values. I understand there may be doubts. So, I will reproduce some statements from the paper and discuss them again: Consider an individual polymeric cylindrical rod which behaves as a torsional spring under external torque. It returns to its original position once the applied torque is removed. We can think of a situation consisting of tightly packed straight cylindrical rods with no buckling. We can twist one of the rods under a localised torque. Then, the twist deformation could stay locally as long as its restoring force is balanced by the frictional force of the surrounding rods. Similarly, we can stretch one rod and it will pop out and stay, like a braille surface, as long as the friction force can balance the restoring force due to stretching. We find the soft hyperelastic rods, with low bending stiffness and high frictional interaction under strain, to be an appropriate system for displaying emergent features such as reprogrammable deformation storage. What I mean here, is that we don’t necessarily need to store the deformation in buckling, which is very space consuming. Assume parallel rubber pillars and we press them together so that they are frictionally locked with each other. You can twist them individually, without taking too much space and the twist will stay, because it’s touching the neighbors which are preventing it from restoring itself. Or increase or decrease length of one of them and that will stay as well. Here’s an example of the kind of rubber pillars I’m talking about. An example of parallel rubber pillars. Image Source And what about miniaturization? Since the effect is again, geometric in nature, that means it is less dependent on the materials we make it with, or the scale at which we make it. The more important thing is that the rods in the network are interconnected in a way that they show hyper-elasticity. Deform like sand (and store memory) or reset itself like an eraser restoring it’s original shape when forces on it are removed. At the end, I must mention some of the drawbacks and difficulties in implementing such a system. Firstly, combining a mechanical system with an electronic system is very hard to manufacture. Secondly, writing the memory of deformation can be time consuming, even if we find optimum ways of reading back the memory of deformation. Finally, it is an all-at-once reset system. It will reset all the memories at once when relaxed.