3 ms·

I'll let him comment on what you guys were talking about but I know of his theory that you are commenting on. But I'm pretty sure you have misunderstood what he

I'll let him comment on what you guys were talking about but I know of his theory that you are commenting on. But I'm pretty sure you have misunderstood what he was talking about in terms of what sparsey does vs his ideas on parts of it (SDRs specifically) being framed as an alternative to quantum computing. You should have linked his direct thoughts on the qunatum computing parts, instead of throwing all of Sparsey under the bus. http://sparsey.com/SDR_and_QC.html http://sparsey.com/SDR_and_QC.html You may have not directly called it quackery but you suggested and alluded to it requiring a bogosity meter and that quacks are attracted to it. It's pretty clear what you meant, to be honest. Sparsey in general though is a separate subject really. What it is doing is capturing all the features it has ever seen and building temporal and spatial relations between those features and then storing those as a code that represents an input in a particular spatial and temporal context. So if I have the words CAT and BAT the A in CAT and the A in BAT will have different codes because of the letter that came before it, likewise with T because A's code is different for both C and B therefore so is T's in both separate cases. So I think that is where the whole quantum part comes into his ideas. Every sample in a sequence has many possible "realities" or codes that can represent it and by representing it in a certain way you are also representing an entire sequence that came before it and a branching point of possible future representations. I have near zero experience with quantum mechanics so I won't comment on how closely it is or isn't to that, because I have no clue. More importantly though, even if it doesn't represent an alternative to mathematically tit for tat exactness for quantum computing, what it can do is reduce computation times dramatically. The idea is that for an equation you could give a sufficiently sized sparsey model a number of inputs. Sparsey naturally represents those inputs as some representation and you can assign the answer to that representation, and you do that with all or as many inputs as possible, each getting its own unique representation. So basically what you are doing is making a giant lookup table of answers. So when you see a set of inputs you know the answer via code lookup rather than calculating it. Sparsey is a bit memory heavy so you would only really want to do that with large hard problems. But that's where the speed comes in, you know instantly the entire state, its past, and its likely conclusion (because it probabilistic), or in fact all of its possible conclusions. It's similar to Q Learning in some ways, in the sense that you are storing many many states you have encountered. Except sparsey doesn't store the states brute force like qlearning. Similar states, spatially, temporally or both, are stored with more code overlap, dissimilar ones have less overlap. But that's the gist of it. He has some results posted. We've talked about getting some sort of public repository up and running for people to play around with and maybe a community forum or something, but that's one of his concerns lately, letting people see the results for themselves rather than taking his word for it. Which, I have one project I need to finish right now, but I will be working on that full time sometime in May.

CodiePetersen 7y ago
Also briefly, doesn't necessarily have to be an equation that the system is learning. Any sequences of samples could be assigned an answer/label and can have overlap with other data types of samples. So an image can be associated with a sound or a word etc, and presenting one or part of the sample could trigger the recall of any of the samples associated with the given sample. So the word tiger triggers the recall of the image, the sound, any context it is found in, etc.
lisper 7y ago
> I'm pretty sure you have misunderstood what he was talking about in terms of what sparsey does That is quite possible. > It's pretty clear what you meant, to be honest. What I meant was exactly what I said: evaluating Sparsey makes a good exercise. I believe it is bogus, but I could be wrong. I am basing my judgement on very sparse (no pun intended) data. > I'll let him comment on what you guys were talking about Here are the relevant quotes from our correspondence: Gerard: I just watched your 2011 "Quantum Conspiracy.." talk and thought you might be interested in an idea I've had concerning quantum theory and how information is represented. The core idea is briefly described in my 2012 paper. "Quantum Computation via Sparse Distributed Representation". I have to related Medium articles as well, "Quantum Computing in the Offing" and "Not Copenhagen, not Everett, a New Interpretation of Quantum Reality". The essential idea is this. To my knowledge, in all quantum theory (QT) and quantum computing (QC) models published to date, prob. amplitudes (PAs) are represented in localist fashion. That is, each PA is represented in its own physically unique memory location (e.g., 64-bits, 32 for real, 32 for imag.) that is physically disjoint from every other PA. In principal, the number of PAs needed to describe a system is of exponential order, so a localist representation requires exponential memory. BUT, if instead, PAs are represented as sparse distributed representations (SDRs), i.e., small subsets of binary units chosen from a much larger field, then they can all exist (be physically stored) in physical (and classical superposition): each PA is just a different subset of the units and the subsets can overlap. Thus, the exponential number of PAs can be represented in sub-exponential (polynomial) memory, Basically, we're leveraging a third fundamental resource, "code space", which is formally orthogonal to the other two, time and (physical) space (i.e., amount of memory). That is, the code space of an SDR coding field is of combinatorial (exponential) order. Me: Thanks for bringing this intriguing work to my attention. You say "the exponential number of PAs can be represented in sub-exponential (polynomial) memory”. This makes it sound as if you claiming that SDRs allow you to achieve quantum supremacy in a classical machine. Am I reading you correctly? Because if that’s what you’re claiming, then you are almost certainly wrong. Since you don’t seem like a quack, it seems more likely that I’m just misinterpreting what you’re saying, so I thought I should get that cleared up before going any further. Gerard: Actually, you are interpreting correctly and thank you for being open to the possibility that I'm not a quack :) [Lots of details about Sparsey elided.] Me: OK, but in the interests of full disclosure you should know that my Bayesian prior on this [not being a quack] is still not very high :-) So let me ask you another “top-level filter” type question: can Sparsey run Schor’s algorithm? If so, have you done it? Gerard: No I've never tried to run Shor's algorithm. [Lots more details elided.] What d'ya think? Me: I think you should try to run Shor’s algorithm, because if you succeed that will be a slam-dunk proof that your claims are correct. Gerard: Well, I don't really understand Shor's algorithm (from wikipedia). I'd have to study it. I never really have so far. I never responded to that because at that point I was pretty convinced he was a quack. It's possible I'm wrong about that. But I'm very confident that he doesn't understand quantum supremacy, and that his claims about it are wrong.

CodiePetersen 7y ago

Also briefly, doesn't necessarily have to be an equation that the system is learning. Any sequences of samples could be assigned an answer/label and can have overlap with other data types of samples. So an image can be associated with a sound or a word etc, and presenting one or part of the sample could trigger the recall of any of the samples associated with the given sample. So the word tiger triggers the recall of the image, the sound, any context it is found in, etc.

lisper 7y ago

> I'm pretty sure you have misunderstood what he was talking about in terms of what sparsey does That is quite possible. > It's pretty clear what you meant, to be honest. What I meant was exactly what I said: evaluating Sparsey makes a good exercise. I believe it is bogus, but I could be wrong. I am basing my judgement on very sparse (no pun intended) data. > I'll let him comment on what you guys were talking about Here are the relevant quotes from our correspondence: Gerard: I just watched your 2011 "Quantum Conspiracy.." talk and thought you might be interested in an idea I've had concerning quantum theory and how information is represented. The core idea is briefly described in my 2012 paper. "Quantum Computation via Sparse Distributed Representation". I have to related Medium articles as well, "Quantum Computing in the Offing" and "Not Copenhagen, not Everett, a New Interpretation of Quantum Reality". The essential idea is this. To my knowledge, in all quantum theory (QT) and quantum computing (QC) models published to date, prob. amplitudes (PAs) are represented in localist fashion. That is, each PA is represented in its own physically unique memory location (e.g., 64-bits, 32 for real, 32 for imag.) that is physically disjoint from every other PA. In principal, the number of PAs needed to describe a system is of exponential order, so a localist representation requires exponential memory. BUT, if instead, PAs are represented as sparse distributed representations (SDRs), i.e., small subsets of binary units chosen from a much larger field, then they can all exist (be physically stored) in physical (and classical superposition): each PA is just a different subset of the units and the subsets can overlap. Thus, the exponential number of PAs can be represented in sub-exponential (polynomial) memory, Basically, we're leveraging a third fundamental resource, "code space", which is formally orthogonal to the other two, time and (physical) space (i.e., amount of memory). That is, the code space of an SDR coding field is of combinatorial (exponential) order. Me: Thanks for bringing this intriguing work to my attention. You say "the exponential number of PAs can be represented in sub-exponential (polynomial) memory”. This makes it sound as if you claiming that SDRs allow you to achieve quantum supremacy in a classical machine. Am I reading you correctly? Because if that’s what you’re claiming, then you are almost certainly wrong. Since you don’t seem like a quack, it seems more likely that I’m just misinterpreting what you’re saying, so I thought I should get that cleared up before going any further. Gerard: Actually, you are interpreting correctly and thank you for being open to the possibility that I'm not a quack :) [Lots of details about Sparsey elided.] Me: OK, but in the interests of full disclosure you should know that my Bayesian prior on this [not being a quack] is still not very high :-) So let me ask you another “top-level filter” type question: can Sparsey run Schor’s algorithm? If so, have you done it? Gerard: No I've never tried to run Shor's algorithm. [Lots more details elided.] What d'ya think? Me: I think you should try to run Shor’s algorithm, because if you succeed that will be a slam-dunk proof that your claims are correct. Gerard: Well, I don't really understand Shor's algorithm (from wikipedia). I'd have to study it. I never really have so far. I never responded to that because at that point I was pretty convinced he was a quack. It's possible I'm wrong about that. But I'm very confident that he doesn't understand quantum supremacy, and that his claims about it are wrong.