What exactly are modern analog computers and how are they better?

[No Wikipedia Cites]

A professor told me that the future of computers was with analog, and that they would be capable of greater power such as more sophisticated artificial intelligence. An attempt to review what the benefits of a modern analog computer would be made things worse in my head.

SO what exactly are modern analog computers (not astrolabes but CPU electric-run) and why are they better? What do they do that is better and how could they even work?

[Keeve]

Originally Posted by sassyfras

modern analog computers (not astrolabes but CPU electric-run)

Sounds like a contradiction to me. The most recent analog computer I can think of is a slide rule.

Then I looked it up on Wikipedia: Electronic analog computers

[Q.E.D.]

They are not better. They are worse. Digital signals are unambiguous, in the absence of severe noise. Even mild to moderate noise is enough to skew an analog system, and when you're doing multiple billions of calculations in your program, those errors are cumulative. Error correction of the sort used in digital systems is not possible, compounding the problem.

[billfish678]

Well, WAG guess here.

The human brain is more analog than digital (maybe) and certainly noisy (hush Bob, I dont care about the army of the 12 monkeys! I am trying to type here), so one could make the arguement that analog might be required for artificial intelligence.

You can simulate analog with digital, but that might still cause problems.

And certain calculations, even done analog, are probably IMO good enough when done analog even though in theory they might not be a precise as digital. A crappy answer with somewhat uncertain precision/accuracy is still better than one where the accuracy/precision is well known but you never get the answer.

[friedo]

Originally Posted by sassyfras

A professor told me that the future of computers was with analog, and that they would be capable of greater power such as more sophisticated artificial intelligence.

Like most professors, this guy has his head in the clouds and is completely disconnected from reality.

We've spent the past 60 years inventing and improving the digital computer. Digital signals have discrete values which can be communicated without degradation and checked for errors.

Digital data can be stored in an unambiguous way, with a known level of precision. The precision of an analog system depends upon its components, and there's no guarantee that two uncalibrated systems will treat a given analog value exactly the same way.

Analog calculations are generally one-way, involving the summing of currents and similar operations. They require massive amounts of hardware to implement complex operations which on digital systems would be issues for software or the OS to manage.

But if you want to see a cool, modern analog computer, this guy build an entire differential analyzer out of Meccano.

[Pushkin]

Originally Posted by sassyfras

A professor told me that the future of computers was with analog

He wasn't mistaking it for Quantum computing by any chance?

[Tyrrell McAllister]

Originally Posted by sassyfras

A professor told me . . .

A professor of what?

[smiling bandit]

Digital computing is moe unambiguous, which is why we use it. However, analog computing has advantages precisely because it is ambiguous. That's its basic advantage, and allows it to do things digital systems can't. There's a non-trivial argument that you can't fundamentally make an intelligent digital AI, because you wind up destroying the thing you're trying to make; it can't think with just on/off states. (the theory is more complex and deeper than I present here, but it's not easily dismissed).

[Tyrrell McAllister]

Could you point to a presentation of this theory that does justice to its complexity and depth?

I'm skeptical that anyone has built an analogue computer that can "do things", in any precise functional sense, that a digital computer cannot.

[billfish678]

Originally Posted by Tyrrell McAllister

Could you point to a presentation of this theory that does justice to its complexity and depth?

I'm skeptical that anyone has built an analogue computer that can "do things", in any precise functional sense, that a digital computer cannot.

Back in the day optical computing was all the rage in theory.

Take an image, many pixels by many pixels. Do a 2-D fourier transform on it. Thats computationally intensive as hell.

A lens "automatically" does that to a whole image at the speed of light. So, in theory, its massively parallel and about as FAST as you can get. And it doesnt generate any heat in the process.

I guess the problem is most computations are not easily/efficiently "transformable" to a 2D fourier transform problem in order to be solved. An then there is the data I/O problem. But again in theory its da bomb.

[Tyrrell McAllister]

I would say that such a computer is not da bomb even in theory if most computations cannot be efficiently formulated in terms of the computer's functions. At best it is da bomb in a very incomplete theory.

[smiling bandit]

Originally Posted by billfish678

A lens "automatically" does that to a whole image at the speed of light. So, in theory, its massively parallel and about as FAST as you can get. And it doesnt generate any heat in the process.

That's basically the principle.

A neuron, for example, doesn't just fire in sequence. They fire in weird patterns, according to their own internal logic and connections, and their nature can change over time. Digital calculations are always done in sequence and each is entirely unconnected to the next.

Now, the problem (among others) here is that, ironically, Digital can't handle ambiguity. AN analog system can be self-correcting. A digital one will rapidly go out of whack, because it can never check its own work properly. It makes a mistake - maybe a tiny one, or it's even not a mistake at all, but just bad data. But that knocks the next calculation out, and the next, and the next. Bam, the system breaks.

Analog systems can take in the whole data set, and errors are adjusted for automatically. Missing data can be assumed. You can adjust errors on digital systems, but then you have to have a whole 'nother system checking for them, and then another to check that, and another. A closely related function is massive parrallelism: analog systems are inherently parrallel-function designs.

Basically, digital is extremely precise but limited in "robustness". Analog is unlimited in robustness but limited in precision.

http://www.mikiko.net/library/weekly...s/aa053198.htm - diasgrees but explains the basic idea. I'm having trouble finding better sources right now because of too much garbage on google, and some older ones are dead links.

[Tyrrell McAllister]

smiling bandit, where are you getting these notions about how digital computers work? Where ever you are getting them, I suggest finding another source. You have been severely misinformed.

Start by learning about error correcting codes.

[billfish678]

Originally Posted by Tyrrell McAllister

I would say that such a computer is not da bomb even in theory if most computations cannot be efficiently formulated in terms of the computer's functions. At best it is da bomb in a very incomplete theory.

Are you here to contribute to the discusion or nitpick?

How many qualifiers does this place require before somebody will let something "slide"?

Geezus.

I am a walking biological 3 D AND temporial transform system. Get back to me when you've got a digital equivalent of me

[smithsb]

Personally, I think Pushkin had the answer - should have been Quantum Computing.

Digital vs. analog. Not perhaps a computer but sound recording and reproduction may be an apt analogy. Digital slices up an analog waveform into smaller and smaller samples to "approximate" the information. Sampling rates and bit depth are factors. An analog recording system may record on tape or other media the actual waveform. Digital reproduction at high rates and depths get close to the original waveform/sound, it can also sound pretty sh**y and lifeless at low rates (crappy MP3s). Analog reproduction gets closer to realism but does come up with the crackles, hiss, and pops. Your favorite LP or tape may develop "noise" but still sounds like music. The CD, if damage/deteriorated, simply won't play. I am aware that many / perhaps most sound reproduction systems may incorporate digital elements (switching amplifiers for example) but analog front to back is still viable.

I can see in a few cases where a system that will accept higher noise levels could be desirable to one that breaks down in spite of error correction algorithms.

[erislover]

Man, and I just read about some new chip that uses fuzzy logic or analog or something. They even created an example chip and its power consumption was much lower. Please, someone has to remember this announcement.

[friedo]

Originally Posted by erislover

Man, and I just read about some new chip that uses fuzzy logic or analog or something. They even created an example chip and its power consumption was much lower. Please, someone has to remember this announcement.

You can do fuzzy logic digitally. In fact, a digital computer can simulate any analog computation (to within a certain degree of accuracy.)

[erislover]

Yeah my post was way too hasty and now I cannot find the reference anymore. I've been searching blogs I usually visit and some tech sites but I cannot find it anymore.

[L. G. Butts, Ph.D.]

Statements of "analog computers are not better" are misleading at best, if not flat out wrong, at least in my opinion. I should probably say "cite?" to these claims, but this is probably not productive.

The term analog computing is pretty vague and covers a lot of ground. I am not a computer scientist so it would be difficult for me to talk generally about the differences and advantages of analog computers. I can however talk about stuff I have done: in grad school I worked on a project to use spectral holography to do signal processing of range/Doppler lidar signals (here is a publication with the initial research). This is absolutely analog processing / computing using optical signals instead of more conventional electronics and it is orders of magnitude higher speed and with orders of magnitude more bandwidth than is possible with current digital technology. More generally, note that a simple lens preforms a 2-D Fourier transform of coherent fields at its focal plane. Take a pixellated image (say a 1000 x 1000 spatial light modulator) and a good laser and you can easily do a 2-D Fourier transform in less than a microsecond. This is equivalent to 10^18 analog multiplies/s. If you were using a standard FFT N log(N) calculation, this rate would be equivalent to 2 teraflops. This is only for a single lens. The problem with most optical processors is that they perform only very specific operations (Fourier Transforms, Convolution / Correlation / Filtering / Pattern Recognition, etc...). But they do it very well, much faster than a digital system could hope to.

Here is an excellent presentation (warning PDF) by a co-student (is this a word?) of mine on a squint compensated RF imaging array system capable of more than 10^17 flops of image processing. Find me a computer that can match it...

[beowulff]

Originally Posted by erislover

Yeah my post was way too hasty and now I cannot find the reference anymore. I've been searching blogs I usually visit and some tech sites but I cannot find it anymore.

Are you thinking of this:
http://www.electronista.com/articles...versity.pcmos/ ?

[RaftPeople]

Originally Posted by smiling bandit

Digital computing is moe unambiguous, which is why we use it. However, analog computing has advantages precisely because it is ambiguous. That's its basic advantage, and allows it to do things digital systems can't. There's a non-trivial argument that you can't fundamentally make an intelligent digital AI, because you wind up destroying the thing you're trying to make; it can't think with just on/off states. (the theory is more complex and deeper than I present here, but it's not easily dismissed).

Do you have a link to this theory or a name, I would be interested in reading their reasoning. It seems to me that a digital system with enough discrete states (for each neuron, for example), would at some point be indistinquishable from a continuous system within the boundaries of the age of the universe, for example.

[Voyager]

Originally Posted by smiling bandit

That's basically the principle.

A neuron, for example, doesn't just fire in sequence. They fire in weird patterns, according to their own internal logic and connections, and their nature can change over time. Digital calculations are always done in sequence and each is entirely unconnected to the next.

There is a big difference between processing an instruction or microoperation and a digital calculation. Digital gates obviously work in parallel. The cite talked about neural networks, which can be implemented digitally also. It has nothing to do with analog computing or even analog circuits.

Now, the problem (among others) here is that, ironically, Digital can't handle ambiguity. AN analog system can be self-correcting. A digital one will rapidly go out of whack, because it can never check its own work properly. It makes a mistake - maybe a tiny one, or it's even not a mistake at all, but just bad data. But that knocks the next calculation out, and the next, and the next. Bam, the system breaks.

Actually, totally wrong. The great thing about digital logic is that each gate is a little amplifier that cleans up the signal. Analog circuits are not like this, and analog circuitry inside a chip is usually very small, tens of components, as opposed to the hundreds of millions inside a microprocessor.

Analog systems can take in the whole data set, and errors are adjusted for automatically. Missing data can be assumed. You can adjust errors on digital systems, but then you have to have a whole 'nother system checking for them, and then another to check that, and another. A closely related function is massive parrallelism: analog systems are inherently parrallel-function designs.

I think this is a case of you mistaking neural nets for analog.

Basically, digital is extremely precise but limited in "robustness". Analog is unlimited in robustness but limited in precision.

That word precision doesn't mean what you think it means. Digital logic is extremely robust. Plus there is a gigantic field of fault tolerance that says how digital systems can recover from any number of errors - assuming you want to pay for enough coding and redundancy. ECCs are just the tip of the iceberg in this area.

http://www.mikiko.net/library/weekly...s/aa053198.htm - diasgrees but explains the basic idea. I'm having trouble finding better sources right now because of too much garbage on google, and some older ones are dead links.

Your source is 11 years old. The lack of more recent links should be a dead giveaway.

[Voyager]

Originally Posted by beowulff

Are you thinking of this:
http://www.electronista.com/articles...versity.pcmos/ ?

That kind of thing - and quantum computing is inherently stochastic, is probably the wave of the future. Luckily it won't happen until long after I retire. I work in hardware testing, which is tough enough when the design is deterministic! This stuff isn't analog in the commonly used sense of the word, though.

[Voyager]

Originally Posted by Tyrrell McAllister

I would say that such a computer is not da bomb even in theory if most computations cannot be efficiently formulated in terms of the computer's functions. At best it is da bomb in a very incomplete theory.

When I was in Bell Labs some people in my center were working on manufacturing issues for optical computing and optical switches. They failed because they never could get cheaper and better than digital switches, but the theory was all there. My old director, who came from Area 11, was an expert on this, and she is a lot smarter than me.

[Voyager]

Originally Posted by smiling bandit

Digital computing is moe unambiguous, which is why we use it. However, analog computing has advantages precisely because it is ambiguous. That's its basic advantage, and allows it to do things digital systems can't. There's a non-trivial argument that you can't fundamentally make an intelligent digital AI, because you wind up destroying the thing you're trying to make; it can't think with just on/off states. (the theory is more complex and deeper than I present here, but it's not easily dismissed).

Not easily dismissed? Just try me.

Somethings are inherently analog, and lots of chips these days have analog blocks which interact with the digital ones by DACs and ADCs. There are also SiPs (system in packages) which have analog and digital dies sitting together in one package. But you can do anything in digital that you can in analog, and I assure you the ambiguity is not an advantage. There is also no way in hell that you can build a practical analog design big enough to do AI.

[Chronos]

It should be noted that there is not a dichotomy between optical computing and digital computing. Most of the optical computing work you hear about is digital, just using photons instead of electrons. Using a lens to do a Fourier transform is a completely different sort of operation from constructing a NAND gate that works on light.

[L. G. Butts, Ph.D.]

Originally Posted by Chronos

It should be noted that there is not a dichotomy between optical computing and digital computing. Most of the optical computing work you hear about is digital, just using photons instead of electrons. Using a lens to do a Fourier transform is a completely different sort of operation from constructing a NAND gate that works on light.

Totally correct. When I think of optical computing, I think of soliton dragging and similar technologies. Optical signal processing is the spacial/spectral frequency based processing done with refraction/diffraction...

[billfish678]

Originally Posted by Voyager

age. There is also no way in hell that you can build a practical analog design big enough to do AI.

I've got a 140 pound 1.0 version sitting right here that does it fine. Though it is rather stinky and ugly. It even runs on SPAM and beer under the right conditions.

[erislover]

Originally Posted by beowulff

Are you thinking of this:
http://www.electronista.com/articles...versity.pcmos/ ?

I am, and it seems to have nothing to do with analog. But thanks!

[bump]

Originally Posted by friedo

Like most professors, this guy has his head in the clouds and is completely disconnected from reality.

Yep... like the professor who taught me computer architecture and had a huge hardon for stack architecture computers, and was convinced that once he worked out the problems with running multiple processes, that they'd take the world by storm.

It's 15 years and counting....

[enipla]

And here I sit. A GIS programmer stunned by this discussion, and my satellite dish is basically down. 80k down and 12k up (I’ve been on the phone for two hours to ‘Rachael’ in India’.

Stunning ideas from what I can manage to read.

Consider what GIS was 20 years ago. You now have it in your phone.

ANALOG computers?

Or is it Analog programming and interpretation? Are we looking for a Boolean field that includes ‘maybe’? Is that it?

[enipla]

Originally Posted by smiling bandit

That's basically the principle.

A neuron, for example, doesn't just fire in sequence. They fire in weird patterns, according to their own internal logic and connections, and their nature can change over time. Digital calculations are always done in sequence and each is entirely unconnected to the next.

Now, the problem (among others) here is that, ironically, Digital can't handle ambiguity. AN analog system can be self-correcting. A digital one will rapidly go out of whack, because it can never check its own work properly. It makes a mistake - maybe a tiny one, or it's even not a mistake at all, but just bad data. But that knocks the next calculation out, and the next, and the next. Bam, the system breaks.

Analog systems can take in the whole data set, and errors are adjusted for automatically. Missing data can be assumed. You can adjust errors on digital systems, but then you have to have a whole 'nother system checking for them, and then another to check that, and another. A closely related function is massive parrallelism: analog systems are inherently parrallel-function designs.

Basically, digital is extremely precise but limited in "robustness". Analog is unlimited in robustness but limited in precision.

http://www.mikiko.net/library/weekly...s/aa053198.htm - disagrees but explains the basic idea. I'm having trouble finding better sources right now because of too much garbage on google, and some older ones are dead links.

I'm going to have to reset the breakers. I do believe you just blew my mind. Very interesting concept.

But it is still data, is it not? It has to be checked and compared. I also think that you are looking at perhaps real time temporal systems that look at the 4th dimension which is time. Time can’t be measured. Not in a computer sense.

[Indistinguishable]

Your mind's been blown by nonsense. As others have pointed out, analog computers are the ones which are susceptible to error-buildup; digital computation is the one which can error-correct. Digital data is far more robust than analog data (if a signal is meant to be either a 0 or a 1, there's rarely any ambiguity as to what to boost it back up to, even if it degrades a little, so to speak. On the other hand, if a signal can vary across a continuous range, then once some small error is introduced, it is generally not possible to determine that such has happened and correct for it).

You're also spouting some nonsense of your own... "Time can't be measured. Not in a computer sense." What does that mean?

[Q.E.D.]

Originally Posted by billfish678

I've got a 140 pound 1.0 version sitting right here that does it fine. Though it is rather stinky and ugly. It even runs on SPAM and beer under the right conditions.

You really shouldn't eat junk mail, you know.

[Der Trihs]

Originally Posted by billfish678

I've got a 140 pound 1.0 version sitting right here that does it fine. Though it is rather stinky and ugly. It even runs on SPAM and beer under the right conditions.

Actually, no. The human brain is heavily digital, when you get right down to it. A neuron fires, or it doesn't. A neurotransmitter is released, or it isn't. That digital aspect is how a living nervous system can function, despite organic sloppiness; the analog uncertainty is pared away into "this neuron has fired". "1", instead of a "0", in other words.

[ultrafilter]

I have seen a device constructed that can decide whether a given number is rational, which is not possible for a digital computer to do. I'm not going to try to give the details, but it's in chapter 33 of this book.

[Chronos]

How do you even specify the number to the device, without making it transparently clear whether it's rational?

[Indistinguishable]

Why say no digital computer can do this? Which is to say, it depends on what exactly this is. It depends on how the input is provided. I mean, if arbitrary real number input is meant to be provided as an analog quantity, then the computation is automatically, at least in part, analog, to the extent that it manipulates that quantity. In that sense, sure, no digital computer can pull this off, but that's trivial.

(Incidentally, in case anyone is curious, the method given in that book is "Shoot a laser into a pinhole at the corner of a square box lined with mirrors, the slope of its direction being the input number. If it ever comes back out of the pinhole, that slope was rational." Of course, this method has zero error-tolerance (as would any computation trying to distinguish the rationals from the irrationals); it depends on the pinhole being exactly one point with 0 width and so forth)

ETA: Chronos beat me to the first point, somewhat

[Frylock]

Originally Posted by Der Trihs

Actually, no. The human brain is heavily digital, when you get right down to it. A neuron fires, or it doesn't. A neurotransmitter is released, or it isn't. That digital aspect is how a living nervous system can function, despite organic sloppiness; the analog uncertainty is pared away into "this neuron has fired". "1", instead of a "0", in other words.

But the threshold at which a neuron fires (i.e., the strenghth of incoming signal required to cause the neuron to fire) changes over time and can vary smoothly.

So perhaps at any given point in time, the brain can be construed as a digital computer. But it seems more accurate to say that the brain is an analogue machine. If there's something to characterizing the brain as digital, it involves the fact that the analogue machine that is the brain functions to create temporary digital machines.

-FrL-

[enipla]

Originally Posted by Indistinguishable

You're also spouting some nonsense of your own... "Time can't be measured. Not in a computer sense." What does that mean?

Because the mind is not able to count more than minutes or hours easily. And keep track of them.

Time is interesting to me in how it influences our decision making process. And yes. It can be coded. We do it every day when we look at our watch.

But I'm getting off subject.

The idea of an analog computer intrigued me. I write code. All results of my code are the basis of analysis of information. It's either 1, 0 or Null.

I can right code that can create ‘maybes’ or ‘perhaps’ based on information.

The idea that there may be direct information besides 1, 0 or null intrigues me.

[Indistinguishable]

Originally Posted by enipla

The idea that there may be direct information besides 1, 0 or null intrigues me.

What does this mean? If you're willing to think of, e.g., integers as naught but strings of 0s, 1s, and "null"s, well, analog data (i.e., a real number) is just a string of 0s, 1s, and "null"s too. (Albeit a very long string; e.g., the k-th bit of the string specifies whether the real is below (0), above (1), or equal to ("null") the k-th rational number. (Incidentally, in the system I have in mind, it would not generally be possible to affirmatively determine of a value that it is "null", though this is probably not the use of the term you had in mind, and so I should perhaps pick a different name for it; "_|_", say).

[billfish678]

Originally Posted by Q.E.D.

You really shouldn't eat junk mail, you know.

It works on the Data In Garbage Out theory

[lazybratsche]

Originally Posted by Der Trihs

Actually, no. The human brain is heavily digital, when you get right down to it. A neuron fires, or it doesn't. A neurotransmitter is released, or it isn't. That digital aspect is how a living nervous system can function, despite organic sloppiness; the analog uncertainty is pared away into "this neuron has fired". "1", instead of a "0", in other words.

I can't speak much to the computer science side of things, but as a biologist I have to strongly object to this statement. It's true that there are some behaviors in (biological) neural networks that are discrete - the neuron fires, or it doesn't. However, there's a lot more than to it than that. What threshold does the neuron fire at? What are the rates of signal conduction within a neuron? What rate does it fire at? How long does it keep firing? How much neurotransmitter does it release? Which neurotransmitters are released, and in which relative quantities? What is the strength of the connection to the next neuron? Then there are numerous modulating signals at every step of the process, and stochastic behaviors throughout.

At the circuit level, sometimes there are nice discrete behaviors. These are well-studied, comparatively, because they're easy to deal with. Ultimately, neural circuits are built from incredibly squishy, analogue things.

[Voyager]

Originally Posted by billfish678

I've got a 140 pound 1.0 version sitting right here that does it fine. Though it is rather stinky and ugly. It even runs on SPAM and beer under the right conditions.

But it is more digital than analog. The signals don't go from one end of the brain to the other - they cascade through neurons, which regenerate the signal just like gates do. They are a lot more complex than simple gates, but cell libraries these days are also. So you are more digital than you think you are.

[Voyager]

Originally Posted by bump

Yep... like the professor who taught me computer architecture and had a huge hardon for stack architecture computers, and was convinced that once he worked out the problems with running multiple processes, that they'd take the world by storm.

It's 15 years and counting....

Hmm. The real problem was that 15 years ago he was already about 15 years behind the times.

[Voyager]

Originally Posted by Frylock

But the threshold at which a neuron fires (i.e., the strenghth of incoming signal required to cause the neuron to fire) changes over time and can vary smoothly.

So perhaps at any given point in time, the brain can be construed as a digital computer. But it seems more accurate to say that the brain is an analogue machine. If there's something to characterizing the brain as digital, it involves the fact that the analogue machine that is the brain functions to create temporary digital machines.

-FrL-

The real distinction between an analog and digital is that the output of an analog block is a smooth function of its inputs, while the output of a digital cell may be controlled by its inputs, but isn't directly proportional to them. That's true no matter how complex the input function is that causes the output to fire. You can (and I think I have for some reason or other) build a cell that only fires when you get a certain number of 1s on the input. You can have as many inputs as you want to get any level of precision. You can even add control inputs to change this. It is still digital because the output pretty much looks the same no matter what the input is.

A while back there was a lot of work on multivalue logic, which used more than 0 and 1 at an input. There was even an IEEE technical committee on this. I think it vanished, since voltages are so low these days that you'd have major problems with noise at the inputs, but IIRC they were inspired by neurons.

[Indistinguishable]

Originally Posted by Voyager

A while back there was a lot of work on multivalue logic, which used more than 0 and 1 at an input.

What kind of work on multivalued logic was this? What I mean is, in a sense, the computer on my desk handles 256-valued logic just fine, only when it operates on such values, we call them "bytes" instead of "bits". What above and beyond that sort of thing was meant by a computer architecture which used multivalued logic (or am I misguided from the start in thinking this was work on computer architectures)?

ETA: Or perhaps you were talking specifically about multivalued logic using real values and continuous functions upon them, rather than just a discrete set ("fuzzy logic" and all that)? I suppose that would make more sense in the context of this thread. In fact, that must be what you meant. Yeah, I'm dumb. Ignore me...

[Voyager]

Originally Posted by enipla

The idea of an analog computer intrigued me. I write code. All results of my code are the basis of analysis of information. It's either 1, 0 or Null.

I can right code that can create ‘maybes’ or ‘perhaps’ based on information.

The idea that there may be direct information besides 1, 0 or null intrigues me.

Check out Fuzzy logic . It can be implemented without analog, but it is mighty useful. For instance, in data mining, not all things fall into neat clusters, and it is useful to classify things as kind of tall or kind of short.

[Voyager]

Originally Posted by Indistinguishable

What kind of work on multivalued logic was this? What I mean is, in a sense, the computer on my desk handles 256-valued logic just fine, only when it operates on such values, we call them "bytes" instead of "bits". What above and beyond that sort of thing was meant by a computer architecture which used multivalued logic (or am I misguided from the start in thinking this was work on computer architectures)?

Or perhaps you meant multivalued logic using real values and continuous functions upon them, rather than just a discrete set ("fuzzy logic" and all that)? I suppose that would make more sense in the context of this thread. In that case, ignore me...

This was in building gates which inherently handled more than two values at the input. I think Intel came out with a memory that did this, but I don't know what happened to it. I don't think it is used today.

The benefit is that you can store two bits of information in the space required for one. The downside, which is the thing that I suspect killed it, is that you have less room for fluctuations of input voltage. This means you have to go slowly, since there is a lot of voltage swing, and that you have to use high voltages, which eats up power. For a 5 volt design, a digital circuit may be a 0 under a volt and a half, a 1 over 3 volts, and guarantee that it never stabilizes in the middle. You could still give each of the 4 values a range of about a volt in multivalue logic, which might work. Today we use 1 volts supplies, so there isn't a lot of margin. If you've ever seen a 1 GHz waveform, it would be obvious why this isn't too useful any more.

My expertise in this area mostly comes from eating dinner with the head of the multivalue logic committee at Computer Society Tech Board meetings when I was on that, so I'm not a good cite. Fuzzy logic is kind of like this, but runs on normal computers.

Multivalue logic was more of a circuit design thing than an architecture thing.

If you want to go the other way, you get into an invention of mine, Base 1 arithmetic, which I invented my first year of grad school when three classes felt the need to teach me binary logic yet again. I published a short summary of it as my first column. Base 1 has several advantages - it is immune to noise, quite fault tolerant, and, if you plug in base 1 to the standard information theory equations, you will find that it is very energy efficient. I would venture you could represent most web pages in base 1 with very little loss of content.

[Voyager]

Originally Posted by lazybratsche

At the circuit level, sometimes there are nice discrete behaviors. These are well-studied, comparatively, because they're easy to deal with. Ultimately, neural circuits are built from incredibly squishy, analogue things.

At the gate level, yes, at the real circuit level, where you have to worry about waveforms and the like, things have gotten pretty messy also. But see my comment above - I'd say neurons are fundamentally digital, while admitting they are a lot more complex than simple gates.