Our technology is different than the one used by Lightmatter and Lightelligence. We currently do not do integrated photonics.
We are starting from the ground up by first building a photonic co-processor that does one function (to be extended) of general interest to AI. the current function is a random projection. We believe that it will be complexified over time as we grow. But our particular focus is on those computations that are difficult to perform in electronics. In particular, our recent focus has been on Transformers/Large Language Models where we think our technology will be a game changer in the training and inference of these gigantic Machine learning models. For the training part, we published two papers at NeurIPS last year in the main conference. We also presented a more detailed one at one of the workshop: https://arxiv.org/abs/2012.06373
We are going to present a poster at HotChips this year about our foray in HPC: https://arxiv.org/abs/2104.14429
Other papers in AI are listed at the LightOn AI research page: https://lair.lighton.ai/
Physically, speaking, the Optical Processing Unit fits into a 2U rack (currently our focus has not been to miniaturize yet, we avoid permature optimization) that is connected to a CPU bus through a PCIe connection.
The technology has been running for the past three years anf half in datacenters for the oldest prototypes (we just decommisioned one a month ago or so) with Machine Learning computation unit tests being performed every ten minutes during that time period.
Let us note that, from the article, this translates into only 3 percent of total electricity generated nationwide. This is in large part due to the inability for wind power to generate a constant power load. From what I recall the nuclear power plants can produce a load of about 100 x 1400 MW = 140,000 MW and this is estimated to be about 20 percent of the elctricity produced nationwide.
Roughly speaking it looks like the nuclear power plants have three times as much installed capacity than solar and produce 6 times as much power to the grid. Since the nuclear power plants are near 100 percent capacity, I wonder what is the reason why wind has roughly a 50 percent production capability: is it maintenance or wind availability ?
I agree. Briefly looking at the description of it, it is some sort of compressed sensing. The differences from traditional CS are minimal in fact, but the scheme is in line with some of the work undertaken in manifold signal processing. The differences are:
- the proposed hash is deterministic, generally in CS, you want to rely on random projections (yet there are some results for deterministic problems) in order to get some sort of universality and by the same token some sort of robustness.
- step 3 and 4, are the most fascinating steps because they are clearly one of the approaches used in manifold signal processing for images. To summarize, in order for pictures to be close to each other on a manifold, you really want to defocus them. I'll put something on my blog on the matter. This is the reason why the has of two images next to each other are close in the "hash" or manifold space.
- for one image, the hash seems to provide 16 measurements (16 bits of the hash result). That would be OK if the initial picture was at the size and color of the picture after step 1 and 2. So in effect, that information is lost. However, in CS you also have "lossy" scheme such as the 1-bit compressed sensing approach (there you retain only the sign of the measurement!, i.e. a little bit like step 6). The reconstruction of these 1-bit pictures are not the original but they are close).
The Wired articles gave the wrong impression. It gave an example of inpainting (the Obama picture) that is NOT compressed sensing. Compressed sensing on the other hand can be applied directly to MRI because the MRI machine actually picks up all the spatial information. It samples randomly in the Fourier domain thereby having access to all the spatial information needed to fully reconstruct an image. It used to be that the reconstruction algorithms were not that good before (they relied on SVD/least square). Candes, Tao, Romberg and Donoho then published papers showing that the reconstruction could be done in a totally different way AND it was exact. With these new reconstruction algorithms something like MRI data is acquired in a much more efficient manner than four years because of compressed sensing.
I have some experience in the process of doing the filtering you are mentioning as I write a blog on compressed sensing (http://nuit-blanche.blogspot.com/), a field that spans a wide range of communities that are, often, not talking to each other nor have the same culture.
I don't think Mendeley or any RSS feed services help much as they both target the same dead tree people who are generally about two years behind on the good stuff. Arxiv is good but not all the communities use arxiv either and so it really is a mix of different tools that yield the most results.
I have a high interest in expanding and automating my current processes and would be very much willing to talk to somebody who would want to spend some time on this.
I also agree with one of the posters, how much do you think people would be willing to pay for this service ?
The small problem with GPGPU is that one aspect of this project is to look at several reconstruction solvers (as they are being called in compressed sensing). Not all of these solvers are clearly amenable to a GPU approach. The idea is really to make it easy for new reconstruction algorithm to be tested on these phase transition maps as opposed to a dedicated effort at re-thinking the algorithm so that it can fit the GPU architecture.
Let me rephrase the question to see if I understand it correctly:
Why do you go for a single piece of hardware when in fact, you could already do this on several GPGPUs and one CPU or use several CPUs on the cloud.?
It could probably be done using these techniques but the answer resides in the fact that if the computation of the phase transition is done very fast for a particular "measurement matrix" (i.e. a particular "sensor") then we could foresee its use in near-real time in instances where the "measurement matrix" would be time dependent (such as the atmosphere....). This would be important in that it would allow to figure out if these new "sensors" can be actually used (as opposed to being mere lab constructions).
For compressed sensing to work when one needs to know in advance that the signal (image or anything else really) is sparse or compressible, Once this is known, one needs two findings that are just plain surprising and are the reason there is so much excitement in this area:
- the signal needs to be acquired "incoherently" and there is much work going on in that area, Most hardware is currently not designed to do this incoherent acquisition (this can be done cheaply in MRI which is why you see some applications there) and much work is currently going into designing new sensors that can performed this new kind of acquisition, Much of them are listed here: http://igorcarron.googlepages.com/compressedsensinghardware
- once it is "incoherently" acquired, then one can reconstruct the original signal using linear programming technique solving an l1 problem, But after four years of intense competition between different groups, we now have faster solvers that are approximately solving the l1 problem without linear programming or the lp (p between 0 and 1) and even L0 problems..
I am Igor Carron, the CEO of the company.
Our technology is different than the one used by Lightmatter and Lightelligence. We currently do not do integrated photonics.
We are starting from the ground up by first building a photonic co-processor that does one function (to be extended) of general interest to AI. the current function is a random projection. We believe that it will be complexified over time as we grow. But our particular focus is on those computations that are difficult to perform in electronics. In particular, our recent focus has been on Transformers/Large Language Models where we think our technology will be a game changer in the training and inference of these gigantic Machine learning models. For the training part, we published two papers at NeurIPS last year in the main conference. We also presented a more detailed one at one of the workshop: https://arxiv.org/abs/2012.06373 We are going to present a poster at HotChips this year about our foray in HPC: https://arxiv.org/abs/2104.14429 Other papers in AI are listed at the LightOn AI research page: https://lair.lighton.ai/
Physically, speaking, the Optical Processing Unit fits into a 2U rack (currently our focus has not been to miniaturize yet, we avoid permature optimization) that is connected to a CPU bus through a PCIe connection.
The technology has been running for the past three years anf half in datacenters for the oldest prototypes (we just decommisioned one a month ago or so) with Machine Learning computation unit tests being performed every ten minutes during that time period.
We are making the technology available for researchers to use on our LightOn Cloud cluster (https://cloud.lighton.ai/what-is-lighton-cloud/) and people can even rent one or several OPUs (LightOn Appliance: https://lighton.ai/lighton-appliance/ ).
I agree, we need to redesign our website.
Igor.