The article talks about switching in a single place, as if by accident. The abstract has no mention of switches at all. Yet, "light-based computers" is something completely dependent on switches, and not much of anything else.
If somebody has access to the actual paper, is there any chance that can lead to switches? Or just multiplexing / demultiplexing?
Update: Thanks fferen. It did demuxing. No work related to switches, except for a "in the future, somebody may use a similar technique to make a switch" on the end (yeah, maybe, using some other dozen techniques not yet discovered, and this one). Demuxing is very important, just that it alone does not lead to practical photonics.
Yeah I was going to say... They've done something which is cool, but really we already knew it was possible, and is not the hard bit of optical computing.
Essentially they made a fancy prism. We need an optical transistor.
It's very difficult to get photons to strongly interact with each other. An optical transistor would revolutionize both classical optical computing and quantum computing.
There are optical switches around. They work by making the material transparent or not depending on an impulse.
The problem is that they are big, slow, monochromatic, hard to build and actually consume more energy than transistors to operate. That's why a breakthrough on that would be interesting.
The actual paper is about realization of an algorithmically-designed optical device rather than optical switching. The device is "just" a computer-designed multiplexer.
Could such a device alter the relative phase of two separate beams of light? This optical device would take as input two beams of light with some arbitrary phase difference, and output the same two beams but with their phases changed to some fixed offset relative to each other.
With such a device an actual switch becomes easy: Make the device change the two inputs to be antiphase and combine them. If one of them is missing, the other passes through unphased (pun intended), but if both are present they cancel each other out; aka you have an xor switch!
Yes, a conventional 2-port MMI will generally give you a pi/2 phase shift of one port relative to another. If you combine the 2 ports again with another MMI, you have built a Mach-Zehnder interferometer with ideally pi phase difference (if you've perfectly matched the length of the 2 waveguides) between the top and bottom signal paths. If you insert a phase shifter in one or both arms you can control the light at the output of the MZ by varying the voltage applied to the phase shifter. You can then modulate the voltage and produce an AM or PM signal at the output of the MZM. This is currently how some commercially available photonic communication IC's send data over the network.
There are limitations on how good the extinction (cancellation) can be based on how well the losses are matched in the respective waveguides.
In this case of this paper, I imagine that the phase relationship will be much more complex and it will highly wavelength dependent.
First, waves still pass through each other when destructively interfering. They temporarily cancel out each others' amplitudes, but not each others' change-in-amplitude, so they continue to propagate.
Second, XOR is not a universal gate (even when combined with NOT). XOR can only sum things together (mod 2). For universal computation you need to be able to multiply (e.g. with an AND gate or a Toffoli gate).
More specifically, their algorithm seems to design linear optical systems (multiplexing,demultiplexing,phase shifting,etc are among them: actually any LT given by a complex matrix of phasors wich is at most unitary can be realized). This matrix can be generalized to the continuous case with a self-adjoint operator but I suppose you get size restrictions to reproduce precisely a complicated operator.
Does this sound like something out of Dirk Gently to anyone else?
> Now the Stanford engineers believe they've broken that bottleneck by inventing what they call an inverse design algorithm.
> It works as the name suggests: the engineers specify what they want the optical circuit to do, and the software provides the details of how to fabricate a silicon structure to perform the task.
I'm quite sure the first novel, Holistic Detective Agency described a computer program that could do that. (In the novel it was of course sold to a governmental body, maybe military.)
This is actually what we do in radiotherapy nowadays, it's even called "inverse treatment planning". Basically a radio-oncologist or a dosimetrist contours (manually, with a pen) organs on every slice of a CT scan and then we just say "We want this much radiation in the tumour, and less than Y units in surrounding healthy tissues".
Treatment planning software then simulates dose distributions from a bunch of possible radiation beam collimations and adjusts the amount of radiation coming out of each radiation "field" to minimize a cost function penalizing overdosing healthy tissue and underdosing the tumour(s).
As a computer graphics professional, this fascinates me. The integrals you describe are remarkably similar to a large variety of solutions we employ for area estimation and integration in general. For example, in a path-tracing paradigm, I would (naively probably) consider radiation beam collimation similar to what we call the "cone-angle" of a particular integration method, particularly WRT calculating illumination response for a given surface, etc. Can you describe in more detail what kind of calculations your treatment planning software does? It sounds fascinatingly similar to the advanced physically based rendering algorithms that are in common use in computer graphics these days. For example, I am guessing that radiation "fields" are akin to "light sources"? If so, I would guess that your planning software is doing all sorts of importance sampling of all these sources across a given domain. Anyway, I was an art major so all my jargon is probably off, but nevertheless I find your post fascinating.
I'm not super familiar with computer graphics, so you'll have to let me know if my description fits what you guys do ;)
I found a youtube video (https://www.youtube.com/watch?v=msX1ypCjkK4) that should give you an idea of what I'm describing actually looks like. Specifically it introduces the concept of a multi-leaf collimator which serves as the main collimating device in modern radiotherapy. The other degree of freedom is the angle of the gantry you see rotating around the patient.
Typically for every gantry angle, the treatment planning software would split up an open field (no collimation) into a bunch of 1x1 cm^2 "beamlets" and would simulate what kind of dose distribution you would get inside the patient from each beamlet (you turn the patient CT into a big 3D grid of voxels to simulate dose in).
You then throw all those dose distributions into an optimiser, and you do what's called a fluence map optimisation which gives you the amount of radiation you want to deliver out of each beamlet. This is the optimisation step I described earlier where the cost function is basically a square difference between the dose in each organ from a given set of beamlet weights and what you want the dose to actually be. Healthy tissue is the limiting factor so you give as much as you can to the tumour while making sure that less than X% of the volume of a nearby organ gets more than Y units of radiation. There's a final step at the end that turns the fluence maps into actual deliverable apertures shaped by the multi leaf collimator.
There's a huge amount of work that goes into the simulation aspect. You can't just model the radiation beam as pure light sources that attenuate in the body via some exponential decay because the high energy photons scatter off electrons which themselves scatter around while depositing energy (radiation dose) away from the point of interaction. The gold standard is Monte Carlo simulations (which is my area of research) since you can model the actual physics of particle transport but in practice most clinics will use a faster engine to generate dose distributions. The faster engines typically superpose a primary component (a pure exponential decay) convolved with a kernel representing the energy that gets deposited away from the point of interaction.
That's probably way more information than you wanted ;)
Wow, that video is really cool! The sliding lock-tumbler mechanism offers a really interesting amount of control over the aperture shaping the beam! Sort of like a brush in photoshop!
Well, an abstract description of processes like these always lends extra heft to what your brain will try to rationalize out of a written description. Especially with computational "decision making" people have a tendency to conjure up ghosts or perhaps a little gremlin at the heart of a nest of wires, watching a TV set, and ruminating over what to do next.
But when you stop and think about what collimation actually is, it's just a manner of focusing a projected beam, either with masking, and obstructing the path of a beam (as with x-rays which are high energey photon beams), or also possibly by bending and focusing a beam, using magnets to direct a beam of charged particles (such as positrons).
So, if working in three dimensions, you might wish to control depth of penetration, but honestly, with X-rays you'll only have so much success, so it's really about how many beams converge upon a region, and the shape they create as they cross over each other, while intersecting, when projected from different angles.
There are a number of ways to approach this strategy, of creating three dimensional shapes, by drawing cross-sections in modern 3D animation programs, like Maya and 3D Studio Max.
The easiest way is to draw a spline or a bezier curve, along one axis, then apply a "lathing" function, which duplicates the spline, rotating about the axis, and then connecting the splines at each control point on the spline/curve. Then you get a crude vase-like shape.
So, take that idea, and apply it to a light source with an articulated aperture. The aperture can create a shadow in the shape of a spline. It might strobe exposures, with small, discrete doses, effectively pixelating or rasterizing the dose with many small exposures, or continuously emit radiation while in motion.
Then, if you attach this beam source to a motorized system, that can rotate the source about an axis on a system of rails, and trigger exposures with different aperture shapes while being positioned around a target at the center of the axis of rotation, hey presto! The software-defined shape has guided the beam, using the same sort of motion control that translates coordinates to a set of motors, as has been done with stop-motion animation cameras in movies for decades!
So, it's like the reverse of a camera, and yes, radiation sources are like flash-bulbs, and you selectively cast shadows, by controlling a gate or shutter, and possibly the shape of the opening in a barrier that stands between the source of the target. (scanline, round dot, square...)
EDIT: As loarake mentions, the actual behavior of a radiation beam is not the same as light. When radiation penetrates a medium, each type of radiation may scatter, reflect, refract differently when interacting with the medium, depending on the material, if it's bone, flesh, metal dental fillings or implanted appliances or something else. Many materials may absorb the radiation and express the interaction by radiating heat, or the radiation will ionize the matter, triggering electrical interactions, and chemical decomposition and reactions. This aspect of radiation is truly the pure random factor (hence monte carlo simulations), the unknowable Schroedinger's cat in a box, but it's real and for every dosage, some radiation will ionize some matter eventually. This along with the conversion to heat is the part that kills tumors, causes burns, and exposes film.
The software took in the thing the user wanted to do as input, and made up plausible excuses for why it should be done. A rationalization engine, as it were.
This is a bit different, more of a constraints optimization problem. If you want a really weird story about that, here's one about the evolved FPGA:
What that article says is incredible, awesome and I wonder if I could replicate it. It probably uses some Xilinx proprietary info about the bitstream though.
Actually approachable math - senior undergrad math major or first year grad can easily understand what's going on. ADMM is actually very old primal dual setup solved via Lagrangian, goes back to 1980s I think.
Good reference - https://web.stanford.edu/~boyd/papers/pdf/admm_slides.pdf
The article itself refers to Boyd's Convex Optimization textbook, which is sort of a rite of passage for anybody doing anything math-related at Stanford these days - even their MS in Management course requires a semester of convex optimization.
The topic is fascinating, but the article doesn't talk to how this "reverse design" algorithm works at all, which I think might be very relevant to this audience. Does anyone have a link to more about the design algorithm and how it works?
An arXiv preprint of the scientific article and copies of our previous work on this topic are available on our research group's website[1]. Just search the page for "design"; the most detailed reference would probably be Jesse Lu's thesis[2].
My first guess would be that it combines a standard "black-box" technique for solving optimization problems (e.g. genetic algorithms, simulated annealing), with an electromagnetic simulation at each iteration. But that is just a wild guess :)
Stupid nature paywall honestly makes me furious. Why doesn't altman or some other yc partner throw some money at nature & get all their articles up on the web for free ? /endrant
> Stupid nature paywall honestly makes me furious. Why doesn't altman or some other yc partner throw some money at nature & get all their articles up on the web for free ? /endrant
I did my PhD work in silicon photonics (in a different lab group and not associated with the authors in the paper) and thought I could chime in with some extra background and why this result is interesting to the silicon photonics community.
First off, silicon photonics has already made its way into several products, mostly active optical cables (a device that directly converts an electrical signal to an optical signal). See, for example, Luxtera/Molex, Acacia, and Kotura/Mellanox. Additionally, many other companies have demoed interesting things at trade shows (e.g. Cisco, Intel, Fujitsu, and others).
In general, the appeal of silicon photonics is that we can fabricate almost all of the components of an optical link on a single chip using the same fabrication tools as what you might find in a standard CMOS fab. Modulators, detectors, switches, filters, and other devices have been demonstrated on a single wafer. Many organizations (ePIXfab, IBM, IME A*STAR, Intel, Freescale, and others) have fabrication processes that have all of these devices right next to each other on a wafer and are capable of 25+ Gb/s data transmit and receive.
Others in the comments have mentioned the lack of switches in the article. Making optical switches in silicon has been demonstrated before, usually with either a Mach-Zehnder interferometer or resonant structure. The phase of light or resonance are most commonly adjusted through the thermo-optic or plasma dispersion effect. I'm at work now, but I can dig up references if anyone is interested later.
This result by Piggot, et. al., is most interesting because it is a unique device geometry for performing a wavelength splitting function. The performance of the device itself isn't particularly impressive relative to other devices with similar functionality that have already been demonstrated [1]. Additionally, the use of an MMI structure for wavelength multiplexing is also not novel [2].
So how does this relate to "light-based computers?" The vision that places like IBM research try to sell is that we will eventually integrate photonics (either monolithically, or flipped in some form) onto our processors and memory chips to enable high-throughput on- and off-chip I/O. This is still likely 10 years away from commercial products. Near-term, look for silicon photonics in your data centers and fiber-optic regional, metro, and long-haul networks. (FTTx one day, but silicon photonics currently can't compete in economics with a DML shoved into a TO can.)
Generally speaking, being able to split based on wavelength lets you transmit data on multiple wavelengths to increase your bandwidth. The flow is to modulate each wavelength individually, mux them together, send them through a single fiber or waveguide, and then demux them on the other side. In a switch, you could imagine switching each wavelength individually and optionally combining them into a single waveguide out of each port of the switch.
This particular device could not be used to make an interferometer. The device has 1 input (call it port 1) and 2 outputs (call them ports 2 and 3). If you input 1550 nm light to port 1, most of it goes to port 2. If you input 1310 nm light to port 1, most of it goes to port 3. This also works backwards: if you input 1550 nm light to port 2 most of it goes to port 1. If you input 1550 nm light to port 3, 10% of it goes to port 1 and the other 90% gets radiated outward as loss (crosstalk is -10 dB). So if you tried to input 1550 nm light to both ports 2 and 3 there won't be much interference at port 1 unless there is a large power imbalance between your two input beams.
Ok, poor question on my part, but a great answer. Thanks! Yes, wavelength-division multiplexing can expand the capacity of a single strand.
I meant, packet switching.
Switching packets of photons would be done by transitioning from transparent to opaque/inverted polarity/shifted wavelength/etc.
I'm way out of my depth here but an interferometer seems like one method that current research is looking at to accomplish that. What do you think looks the most promising?
Yes, an interferometer is certainly a very common method to perform switching. What ultimately gets used will depend on the technology/material system. Silica-on-silicon and silicon photonics-based switches will likely use Mach-Zehnder interferometers. MEMS switches currently use movable mirrors or gratings.
There are several different frequencies of infrared. Unless your chip is getting too hot (in the 400°C or more), the best frequencies do not mix with thermal radiation.
The target audience is TV and print media. Watch, this will be on local Bay Area stations within 72 hours to fill 2.5 minutes of the evening broadcast. Most news is targeted at people with a 6th grade understanding of the world.
If somebody has access to the actual paper, is there any chance that can lead to switches? Or just multiplexing / demultiplexing?
Update: Thanks fferen. It did demuxing. No work related to switches, except for a "in the future, somebody may use a similar technique to make a switch" on the end (yeah, maybe, using some other dozen techniques not yet discovered, and this one). Demuxing is very important, just that it alone does not lead to practical photonics.