Phil Pilgrim (PhiPi) said...
...·Who can really cut through Dr. Jim's vagueness without more detail? I'm just trying to make sense of what he said in terms of something that could work. But neither of us can know for sure what he means without summoning ESP at this point.
Dear Phil, (and all),
if Dr.Jim were someone else and he posted a vague question with no code here on the forum, what would we do? Right, we'd not speculate (much), nor summon (much) ESP - we'd politely ask for code and more info so we could understand and be of help.· That's always been the motive of the forum AFAICT.
It's like these guys are asking for our help - but without really asking AND without code or clarity.·· We tried to guide them. Why are we·putting up such a breach of protocol and etiquette- and more over,·falling into speculation? Look what that altering of forum course did: we started getting annoyed, then mad, then made fun, then got sarcastic - and started snipping at each other (some).
Have we temporarily lost our collective forum wits ?· --- our posts in this Dr.Jim and Mr. Mallred threads are like we are dealing with an alcoholic or a drug addict. Such an individual is unpredictable - but the sober, straight folks can mistakenly assume the problem-laden individual is 'really ok' and respond sanely... with the consequence that the insanity repeats (and repeats ...) Of course, I'm not saying Dr.Jim or Mr.Mallred have such issues personally - but we are NOT responding as we normally do.· We're stuck in a goofy loop.
Why?
Of course, we are interested in the topic --- but wouldn't it be fairer to US HERE if they had said:
" Hey you all ... this is what we are trying to do, but aren't there yet, can you offer advice? "
They did almost the opposite - and we called them on it.· I am glad to see that we still have·serious discernment here folks.
cheers
- Howard
PS. Phil I hope you get those filters working ... then at least we have real claim·here to some A.I. (as in Artful Intelligence [noparse]:)[/noparse])
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Post Edited (CounterRotatingProps) : 8/23/2009 1:31:20 AM GMT
The 300 Hz ... 3.4 kHz was the *old* frequency range here in Germany. It was good enough for voice recognition by humans.
Maybe even less filters could do: 300 .. 500 Hz; 500 ... 1000 Hz, 1 ... 2 kHz and 2 .. 4 kHz (almost every octave).
But Phil will report his results.
Nick
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Never use force, just go for a bigger hammer!
The DIY Digital-Readout for mills, lathes etc.: YADRO
The two proposing AI or MI using a different approach is not without merit. Having scanned the various posts/replies to their topics and the fragments of data related to their project and implementation; I surmise that:
1. Dr. Jim is 65+ and has recently retired
2. He worked in the aerispace field for a time and was involved in telemetry
3. He hs several patents related to spacial and imaging technologies
4. He has designed a multi-Propeller chip system intending to emulate AI in an Android sense
5. Pure compute cycles (ie: horse-power) is not the only aspect of an Androids implementation (the choreography of the input sensors, movements and 'learned' activities is)
6. Propellers have no interrupt latency making them ideal for (near) real-time systems
7. Their design is looslely based on
7a. Up to 4 Propellers (32 COGs) [noparse][[/noparse]at 20 MIPS/COG; capable of running either independantly, co-operatively or a combination - that's awesome!]
7b. They are using 2x Bi-directional serial channels between each Propeller
7c. They have a common 8 bit data bus between each Propeller to their main (fast) storage
7d. Mass or (near) line storage is provided via a USB memory stick (if I recall seeing somewhere)
8. The KISS Operating System - (is not an Operating System in the traditional sense of one that manages a system resources and access to same by applications) is:
8a. A task scheduler and command parses/issuer that monitors and controls the other aspects of the "system"
8b. Contains only 16 "instructions".
Let's look at 8b first and ask "What does an Android do?" - In an ideal implementation, it would emulate or mimic Human characteristics in a way that would give an illusion it was pseudo-human - of course with rules and restrictions that inhibit it from damaging anything (ie: the 3 Laws &c.). So how does one accomplish this sort of behaviour? Examine the learning of children as an example; they start by learning simple words, phrases, gestures and so forth. Mistakes are made and re-inforcement and/or corrective actions and examples are provided by more experienced teachers and ultimately, the behaviour becomes what we classify as "normal", traditional or expected and acceptable.
Take the programming of a robot as another example; One can write code to provide a set of primitives and the robot will execute these in their defined sequence. Add various sensors and additional code to allow for a "learning change" to occur as the robot executs its instructions. Example, if you bump into something, use your range sensors to determine the distance and slow dows, change direction or whatever. The idea is quite straight forward.
Getting back to 8b; the main processor of suce an Android, only needs basic "instructons" that form the core of the expected functions inbuilt into such a machine.
a. Run (a pre-defined instruction sequence; or other code)
b. Look (for something)
c. Move (foerward, backward, left, right [noparse][[/noparse]with possible interpolated axis control influenced by sensor input either stored or real-time)
d. Stop
e. Ask (ask a question or request input of some kind related to the current task)
f. Say (something ie: send feedback to the outside of its 'world')
g. etc
The gist of this is that the central processor dispatches these operations to the other COG's, running on other Propellers that are wired and coded to perform the needed tasks. For example, having 2 serial and one 8 bit bus (as described above) allows at least 16 I/O pins free for each of the 4 Development boards - a total of 65 I/O pins that can be used to read/write various data to any type of supportable I/O device. Combined with the ability to transfer the sensory data either "read" or stored between all COG's allows for some fairly high speed processing - the type needed to provide an emulated behaviour for a 1st generation Android. I would not expect any such machine to necessarily play in an ice-hocky game but for some rudimentary tasks such as getting a cold beer or mowing a lawn...
The reference to speech recognition is also not without merit, the removal of "carrier" freqencies is akin to what is done with AM radio transmissions to create sideband frequency transmissions; this are far more immune to EM interference as an example. To apply similar logic to voice communications sounds (no pun intended) quite reasonable - all one needs to do is distill the inflective utterences of "words" without relying on the underlying "carrier/pitch/IM distortions &c.) and one could indeed build better algorythims than those of the past - all using similar (and old) ideas.
For interpretation, consider English communications - spoken or written (apologies to other language speaking people but English is the only Human language I know). Many words in a sentence are or can be considered "noise" words in many cases however, English as an interesting example of complexity has both different meaning for the same words, the same pronounciation for words that may be spelt differently or different meaning for same words depending on how that are pronounced.
This provides a challange for mechanisms that need to deal with them and ensure that the intent is communicated correctly. If thee two have found a reliabe way to distill and interpret these complexities with near 100% success then I for one would be impressed. I too woud like to hear from these designers details or specifics about their project; Not their IP of course but a carrot or two to provide some meaningful incentive to trust their claims.
Anyway, enough for now but in closin...
Does anyone remember a BASIC (game) program called "Animal" from the early 80's? If not, this was a "learning program" - with respect to the programmer (I'll see if I can find a copy and give credit as due).
In a nutshell, this program started by knowing of a fish and a bird. It asked if you were thinking of an animal. You responded "No" it asked you to do so. You responded "Yes" and it tries to guess what it is, using a logical deduction based on questions, provided by you as the game progresses. For example: the first question it asks is "Can it (the animal you are thing of) fly?"
If you answer "yes", it asks "Is it a bird"; to which you may respond "yes"; if yes, then it asks you to think of something elee. If no, then it asks you what it was (the animal you thought of) then it asks you to type a question that can be used to distinguish between a bird and the animal you thought of.
Over the course of several hours, memory/storage limited this can become quite fancinating and a bit of a challange. What does get interesting is to apply out-of-the-box thinking to what you are thinging about. As long as the dialogue used is relevant to your subjects, it can become quite humerous.
Anyway, the point here is the "learning" process and ability to step through a logical deductive sequence to arrive at an appropriate "answer". This was back in the 80's and I've never forgotton how clever this little program was - I'll see if I can find it - would make a great demo for the Propeller and FemtoBasic
he1957: I like this "The idea is quite straight forward"
If I understand correctly AI (or MI) in it's many interpretations and robotics has occupied the minds of very many smart engineers, mathematicians, computer scientists from all around the world for many decades. Many of whom have assumed "The idea is quite straight forward". The net result of all this effort is still not very impressive in terms of intelligence and learning.
Personally I'm going to invest my money in Morpheus and the Largos OS for my AI project.
Largos is mostly an unknown at this stage but Morpheus is a useful cost effective platform anyway.
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
The issues for AI really distill to storage capacity and the mechanisms needed/used to cross reference use{ful,less} data to provide for decision code that gives suitable outcomes for the given "condition". A matter of "learning" and being able to x-ref and respond in an acceptable and appropriate way. [noparse][[/noparse]Store everything but {only}] Remember what's important or useful when needed; forget about it until a "trigger" recalls something else that should also count in the decision for the current - how, without the capacity?
Morse Code - too much of it will make you "dotty"
I bet the typos in my previous post had been:
1. Unimportant
2. Annoying
3. Un-noticed
4. Filled in or corrected automatically because you "understood" the context
5. Re-Read,. verified as having errors then...
Meaninful extrapolation of data is something computers can't do without the storage capacity available to consider all the required possibilities. Regular or extended expressions can generate oodles of patterns but without a reference to determine what is useful in context is wastefull. Time taken to do this does not really matter unless you can't wait for an answer - either right, wrong or unknown. Impatience is why most development efforts followed the "clock speed" path - doesn't mean diddly compared to parallel processing. Yes, I have an X MB of cache multiplied by a Y SMP environment - guess what, an external device has just updated your database; all your fancy caches are now stale - go refresh!
How many words begin with ""W", end with a period and contain between 2 and 9 letters? Does it matter - why or why not?
BTW: I'm not suggesting investing any money in anything; I'm just being open-minded that someone may have found a "better way" of doing things we thought we were already doing; not to imply that what we already do is the best way to do it - that's all.
Harry, the reason we don't think MIT have discovered better way of doing anything is that what little they have released about their methods (and that is VERY LITTLE) is, according to quite a bit of existing research by others, demonstrably wrong.
First, they released a video extolling "Dr. Jim"'s theory of how neurons work, which is simply wrong. Neurons don't work the way he says, even accounting for the fact that he might have made up his own lingo and be using words differently than other researchers. Neurons are not "on or off." There is no reasonable interpretation of neural function that can be reconciled with the ideas in that video; it is not avant-garde, it is simply wrong, and if the claim is that it will work because it is mirroring the example of biology, well that's wrong too because it's not mirroring the example of biology.
Then in this more recent example we get a description of how speech will be recognized, which starts by throwing away most of the information which people have known is there since roughly 1940. Again, this is not "new" or "better," it is simply wrong. There is no way to interpret that blog post even assuming very liberal definitions for things like "carrier" that is at all consistent with what is known about the structure of speech. While one might say MIT are inventing a new wheel because they think the real scientists are doing it wrong, it's kind of hard to justify that when we know for a fact that the human ear also does spectral separation.
I have met some of the people who are doing serious work on AI. There is FAR more to it than pattern recognition, and it is an insult to a lot of people who have literally dedicated their lives to the task to suggest they are too stupid to figure out that stale caches are the problem. (In fact, your statement of that suggests that you don't have a clue how caches work; there's a lot of fancy math involved in making them transparent to the very kind of thing you describe.)
A single neuron is far more complex than any computer, and it still isn't known how they produce conscious behaviour when connected together. It has even been suggested that processing is going on in the microtubules at the quantum level (Hameroff and Penrose) and that this produces consciousness, but most researchers think that this is rather far-fetched.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Leon, I can't quite agree with that either. There's nothing very mysterious about why neurons fire, and there is quite a bit of mystery about how and why they grow new dendrites and synapses. We have some pretty good ideas as to the nature of those algorithms, but nothing we could solidly duplicate.
As for quantum processing, "far-fetched" is about the kindest thing I can say about that sort of thing. I do think that plain old thermal noise has a lot to do with it, possibly in a way that is necessary rather than just murking up otherwise predictable results. The the quantum silliness harks back to the notion that the brain is so powerful and unpredictable that it can't possibly be something simple enough to emulate with a mere computer. I do believe that's poppycock; there are ten billion or so neurons in a human brain, each with thousands of interconnections. Even if an individual neuron is a fairly straightforward thing to emulate (which I think it ultimately will be) that's a lot of complexity compared to the computers we've built so far.
It's known how they fire, of course, but things are much more complex than that, and there is a lot going on inside the individual neuron that isn't understood.
I don't see the relevance of thermal noise to communication between neurons, the effect would be negligible.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
I think the phrase "Remember what's important or useful" perhaps contains the difficult part. Yes one may have the storage and yes one may have the speed. The hard part of what is generally called intelligence is in the "important" and "useful". where does that judgment come from?
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
With apologies, the references were not intended to insult anyone but appear to have done so but I don't see any implication of anyone being "stupid".
References were purely intended to illustrate the _time_ needed for a CPU to flush cache (1st, 2nd or Nth level) and access (any) external [noparse][[/noparse]bus or I/O controller based] memory, multiplied by the number of CPU's on-line in an SMP (Symmetric Multi-Processor) environment to maintain [noparse][[/noparse]cache] coherency; should this be necessary.
WRT "pattern-matching"; this was perhaps an over-simplified example. To expand; the reference was WRT distilling of data and patterns representing 'sentences/commands or whatever' that can then be decoded/replayed/looked-up using appropriate methods.
If one was to encode the "simplified" aural digitisation of "Go to the refrigerator and get a cold drink" into a sequence of a minimal set of bits (with an attached EDC/CRC - or whatever) - it would be, could be or definitely should be unique - if not the encoding mechanism should compare what it encoded with what it "knows" and if it reveals an existing pattern for a (perhaps) different "thing" then it should seek clarification of what constitutes differences. Better that than "go drive the car into the river" because it "thought" it understood.
Finally, the simplistic cache reference was allowing an "external 'device'" to update {memory} - a common resource in this context which of course would have to use either a DMA mechanism; requiring some form of "bus mastership" ability, or, another form of access; controlled by some form of "handler". In either case, any CPU (SMP or other) with a cache mechanism would have to be advised their references to data for addresses affected by the "DMA" cycle would now be invalid [noparse][[/noparse]if any].
This could include a Propeller if they were processing real (or near) time data related to environment and or activity - physical or otherwsie; using data obtained from a "common" source.
Leon -- behold Alopex: http://en.wikipedia.org/wiki/ALOPEX .· The "stochastic process" mentioned in Wiki was in fact first proposed by Erich Harth to be thermal noise (I have his original article from the journal SCIENCE thumbtacked to my wall behind my computer).· It is alopex which convinced me that machine intelligence was probably possible in the mid 1990's after a decade or so of doubt.· Alopex remains more of a pattern recognizer than an intelligence, but it has the curious quality that, without being specifically told to, it tends to make the same kind of mistakes we do.
Harry -- Again, you vastly oversimplify something that has been studied to death.· The time needed for a modern CPU to flush cache is generally negligible because (1) the cache is in small bits, the whole thing isn't flushed and refreshed at once, (2) memory hardware is optimized for block read and write, and most important (3) in a properly designed system cache refresh is usually performed via read-ahead and write-behind while other bits of the cache are in use by the CPU.· This has all been standard in PC's for at least 15 years.
As for aural digitization, we have known since the 1940's both that the "encoding mechanism" for phonemes is a time-variant spectrum, and when you throw away or never extract the spectral information, as MIT proposes, you are not "encoding" anything.· Speech recognition is not actually a total mystery, and we know how the basics work; MIT proposes to ignore all this and take a path which we frankly know doesn't work.· You cannot compare information you have thrown away to anything.
We have also had some time to learn to build massively parallel computers: http://en.wikipedia.org/wiki/Connection_Machine· An interesting fact about the early Connection Machines was that their individual processors were too simple to have caches.· Now considering what these machines are used for and the progress made in the last 25 years, do you really think the people spending millions of dollars to build these things are so stupid that they are going to be tripped up by the equivalent of "OH NOES I GOTZ 2 FLUSH MAH CASH?"
It comes from learning (and needing or wanting) the difference. How much trivia can you recall - given something you find interesting (ie: important or useful) versus "Facts and Figures of 800 BC" you were forced to sit through by your Library Teacher when you were 9?
When you started learning, primitives were set; as were doctrines, diciplines and right and wrong. Being able to determine right and wrong goes a long way; Doing something that is "wrong", fun and worth the subsequent "punishment" is a good risk and worth more if you don't get caught and suffer the punishment - right?
All references are based on the "dance" (choreography) of the actions of determining all avalable stored and real-time data that is used in choosing what to do (be it right or wrong) [noparse][[/noparse]this of course forms the basis of an AI machine - it is not allowed to do wrong because Humans say so].
However without the continual storing of "experiences" [noparse][[/noparse]input data] and filing for "use{ful,less}" recall later, you (not personally but any-critter) are less capable or more likely to make a "bad call" - especially as the details surrounding the "call" get more complex. This is sometimes defined as "experience" - as they say "it comes with age" - why? I've seen people make the same mistakes multiple times for similar reasons - how come? We are supposed to learn from mistakes - right?
Walk down the street on a nice Summer day - a smell triggers a memory - long since forgotten (buried or whatever) - why yes but it's still there in _every detail_ - why and how?
It's the simple things that'll get ya - every time. Some people have become so impressed by their own "intelligence" they sometimes forget their place. One of the Forum participants has a nice signature line - simplicity.
What reasons would there be here for not using simplistic references?
Thermal noise has a negligible effect on brain neurons. If it did have an effect they couldn't work properly. Alopex has nothing to do with real neurons.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Alopex was based on real neurons, specifically the architecture of the Thalamus. And it isn't the only algorithm in this family that requires a random input. Your assertion that noise has nothing to do with how neurons work, right after you asserted that we don't know how neurons work, strikes me as more than a little strange.
The thalamus is part of the primitive limbic system, and isn't involved directly with higher cognitive functions like long-term memory. Thermal noise isn't relevant to inter-neuron communication; there are other sources of noise that are relevant. It's the internal workings of the neuron that are still something of a mystery, not how they communicate with each other.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Well actually the Harth/Unnikrishinan/Pandya article in Science asserts (and is not alone in this) that the pathways through older parts of the brain play a major role in perception. And nobody said anything about noise being important for inter-neuron communication; what I said was that it might be a critical part of the algorithm the neuron is performing which enables it to do its duties.
And having lived long enough to experience what age starts to do to one's brain, I would advise that cell death and noise are probably major parts of the process whether they're necessary or not.
The thalamus receives sensory data, and connects to higher cortical areas. Selective attention is one thing it controls. It's also important in the mediation of fear and anger.
Is there any evidence that neurons actually use noise in their operation?
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
The Harth et al article claims alopex as a model for how the thalamus works, based on then-known fiber wiring. Harth claims that the working of the visual cortex is also very similar, although it is much harder to trace those pathways for verification I believe he based that on observed likeness of function. The noise component was introduced as a necessary input to get the computer models of the neuronal path to function; the alopex algorithm can be competely implemented by the known behaviors of neurons wired as they really are in the limbic system. In this case the stochastic input is not a random destructor of the signal, but a necessary part of the algorithm that allows it to avoid getting stuck in local minima as it sharpens/optimizes. Alopex is not the only optimization/pattern recognition program to use the technique. Harth is very explicit that the thalamus is considerably more than a relay station between the senses and the cortex; it does significant pre-processing and probably has a major role in conscious perception. The stochastic input does mean that the output of an alopex machine is unpredictable in its particular short-term response, but given known programming it will settle into a predictable range of long-term responses.
I know nothing of brain function. Bearing in mind that mine barely functions as it is. But I do remember an explanation that as the brain learns it's not just that signal strengths or rates of firing between neurons that changes but that they are actually growing new synapses forging new links between each other.
This sounds like some kind of chemotaxis to me en.wikipedia.org/wiki/Chemotaxis all be it if the neurons are not actually moving, they are reaching out in some direction.
Which poses the question, how does a neuron know in which way to grow its synapses? What is the chemical gradient or chemotactic signaling they are following? And how did that come to be there?
When I here people suggesting that a neuron is "a simple thing" that can be modeled in a computer algorithm and many instances of it combined into a neural net which will then model a brain, I never here any mention of this chemotactic growth phenomena which seems to already be quite a complicated thing.
Looks to me like a single neuron has a lot of complicated, even "smart" behavior that is overlooked. It's at least as smart as that bacteria swimming around searching for food and avoiding harmful environments.
Still, what do I know.
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
Long-term memory does involve the brain rewiring itself by making new neuronal connections. It must be controlled by something inside the neuron; as I said previously they are very complex devices and aren't simple computing elements.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
When you said that the morse tappings were "used to drive his wife crazy" did you mean that he did it innocently and it upset her, or, it WAS USED to drive her crazy. The last Mrs Seckshund went that way all on her own accord, I think,·as I never learnt morse (only a G8).
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Style and grace : Nil point
Comments
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Propeller Wiki: Share the coolness!
Chat in real time with other Propellerheads on IRC #propeller @ freenode.net
Safety Tip: Life is as good as YOU think it is!
if Dr.Jim were someone else and he posted a vague question with no code here on the forum, what would we do? Right, we'd not speculate (much), nor summon (much) ESP - we'd politely ask for code and more info so we could understand and be of help.· That's always been the motive of the forum AFAICT.
It's like these guys are asking for our help - but without really asking AND without code or clarity.·· We tried to guide them. Why are we·putting up such a breach of protocol and etiquette - and more over,·falling into speculation? Look what that altering of forum course did: we started getting annoyed, then mad, then made fun, then got sarcastic - and started snipping at each other (some).
Have we temporarily lost our collective forum wits ?· --- our posts in this Dr.Jim and Mr. Mallred threads are like we are dealing with an alcoholic or a drug addict. Such an individual is unpredictable - but the sober, straight folks can mistakenly assume the problem-laden individual is 'really ok' and respond sanely... with the consequence that the insanity repeats (and repeats ...) Of course, I'm not saying Dr.Jim or Mr.Mallred have such issues personally - but we are NOT responding as we normally do.· We're stuck in a goofy loop.
Why?
Of course, we are interested in the topic --- but wouldn't it be fairer to US HERE if they had said:
" Hey you all ... this is what we are trying to do, but aren't there yet, can you offer advice? "
They did almost the opposite - and we called them on it.· I am glad to see that we still have·serious discernment here folks.
cheers
- Howard
PS. Phil I hope you get those filters working ... then at least we have real claim·here to some A.I. (as in Artful Intelligence [noparse]:)[/noparse])
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Post Edited (CounterRotatingProps) : 8/23/2009 1:31:20 AM GMT
The 300 Hz ... 3.4 kHz was the *old* frequency range here in Germany. It was good enough for voice recognition by humans.
Maybe even less filters could do: 300 .. 500 Hz; 500 ... 1000 Hz, 1 ... 2 kHz and 2 .. 4 kHz (almost every octave).
But Phil will report his results.
Nick
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Never use force, just go for a bigger hammer!
The DIY Digital-Readout for mills, lathes etc.:
YADRO
They could just hum the morse and even a low mips microcontroller would have no problem
processing the information
-I like the name KISS OS, rolls off the tongue well.
hmm hmmmm hmm hmm - hmmmm hmmmm hmmmm - hmm hmmmm hmm hmm
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
- Some mornings I wake up cranky.....but usually I just let him sleep in -
Post Edited (HollyMinkowski) : 8/23/2009 8:26:43 AM GMT
1. Dr. Jim is 65+ and has recently retired
2. He worked in the aerispace field for a time and was involved in telemetry
3. He hs several patents related to spacial and imaging technologies
4. He has designed a multi-Propeller chip system intending to emulate AI in an Android sense
5. Pure compute cycles (ie: horse-power) is not the only aspect of an Androids implementation (the choreography of the input sensors, movements and 'learned' activities is)
6. Propellers have no interrupt latency making them ideal for (near) real-time systems
7. Their design is looslely based on
7a. Up to 4 Propellers (32 COGs) [noparse][[/noparse]at 20 MIPS/COG; capable of running either independantly, co-operatively or a combination - that's awesome!]
7b. They are using 2x Bi-directional serial channels between each Propeller
7c. They have a common 8 bit data bus between each Propeller to their main (fast) storage
7d. Mass or (near) line storage is provided via a USB memory stick (if I recall seeing somewhere)
8. The KISS Operating System - (is not an Operating System in the traditional sense of one that manages a system resources and access to same by applications) is:
8a. A task scheduler and command parses/issuer that monitors and controls the other aspects of the "system"
8b. Contains only 16 "instructions".
Let's look at 8b first and ask "What does an Android do?" - In an ideal implementation, it would emulate or mimic Human characteristics in a way that would give an illusion it was pseudo-human - of course with rules and restrictions that inhibit it from damaging anything (ie: the 3 Laws &c.). So how does one accomplish this sort of behaviour? Examine the learning of children as an example; they start by learning simple words, phrases, gestures and so forth. Mistakes are made and re-inforcement and/or corrective actions and examples are provided by more experienced teachers and ultimately, the behaviour becomes what we classify as "normal", traditional or expected and acceptable.
Take the programming of a robot as another example; One can write code to provide a set of primitives and the robot will execute these in their defined sequence. Add various sensors and additional code to allow for a "learning change" to occur as the robot executs its instructions. Example, if you bump into something, use your range sensors to determine the distance and slow dows, change direction or whatever. The idea is quite straight forward.
Getting back to 8b; the main processor of suce an Android, only needs basic "instructons" that form the core of the expected functions inbuilt into such a machine.
a. Run (a pre-defined instruction sequence; or other code)
b. Look (for something)
c. Move (foerward, backward, left, right [noparse][[/noparse]with possible interpolated axis control influenced by sensor input either stored or real-time)
d. Stop
e. Ask (ask a question or request input of some kind related to the current task)
f. Say (something ie: send feedback to the outside of its 'world')
g. etc
The gist of this is that the central processor dispatches these operations to the other COG's, running on other Propellers that are wired and coded to perform the needed tasks. For example, having 2 serial and one 8 bit bus (as described above) allows at least 16 I/O pins free for each of the 4 Development boards - a total of 65 I/O pins that can be used to read/write various data to any type of supportable I/O device. Combined with the ability to transfer the sensory data either "read" or stored between all COG's allows for some fairly high speed processing - the type needed to provide an emulated behaviour for a 1st generation Android. I would not expect any such machine to necessarily play in an ice-hocky game but for some rudimentary tasks such as getting a cold beer or mowing a lawn...
The reference to speech recognition is also not without merit, the removal of "carrier" freqencies is akin to what is done with AM radio transmissions to create sideband frequency transmissions; this are far more immune to EM interference as an example. To apply similar logic to voice communications sounds (no pun intended) quite reasonable - all one needs to do is distill the inflective utterences of "words" without relying on the underlying "carrier/pitch/IM distortions &c.) and one could indeed build better algorythims than those of the past - all using similar (and old) ideas.
For interpretation, consider English communications - spoken or written (apologies to other language speaking people but English is the only Human language I know). Many words in a sentence are or can be considered "noise" words in many cases however, English as an interesting example of complexity has both different meaning for the same words, the same pronounciation for words that may be spelt differently or different meaning for same words depending on how that are pronounced.
This provides a challange for mechanisms that need to deal with them and ensure that the intent is communicated correctly. If thee two have found a reliabe way to distill and interpret these complexities with near 100% success then I for one would be impressed. I too woud like to hear from these designers details or specifics about their project; Not their IP of course but a carrot or two to provide some meaningful incentive to trust their claims.
Anyway, enough for now but in closin...
Does anyone remember a BASIC (game) program called "Animal" from the early 80's? If not, this was a "learning program" - with respect to the programmer (I'll see if I can find a copy and give credit as due).
In a nutshell, this program started by knowing of a fish and a bird. It asked if you were thinking of an animal. You responded "No" it asked you to do so. You responded "Yes" and it tries to guess what it is, using a logical deduction based on questions, provided by you as the game progresses. For example: the first question it asks is "Can it (the animal you are thing of) fly?"
If you answer "yes", it asks "Is it a bird"; to which you may respond "yes"; if yes, then it asks you to think of something elee. If no, then it asks you what it was (the animal you thought of) then it asks you to type a question that can be used to distinguish between a bird and the animal you thought of.
Over the course of several hours, memory/storage limited this can become quite fancinating and a bit of a challange. What does get interesting is to apply out-of-the-box thinking to what you are thinging about. As long as the dialogue used is relevant to your subjects, it can become quite humerous.
Anyway, the point here is the "learning" process and ability to step through a logical deductive sequence to arrive at an appropriate "answer". This was back in the 80's and I've never forgotton how clever this little program was - I'll see if I can find it - would make a great demo for the Propeller and FemtoBasic
Cheers,
HarryE.
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
If I understand correctly AI (or MI) in it's many interpretations and robotics has occupied the minds of very many smart engineers, mathematicians, computer scientists from all around the world for many decades. Many of whom have assumed "The idea is quite straight forward". The net result of all this effort is still not very impressive in terms of intelligence and learning.
Personally I'm going to invest my money in Morpheus and the Largos OS for my AI project.
Largos is mostly an unknown at this stage but Morpheus is a useful cost effective platform anyway.
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
OMG!
Imagine what he might have been sending!?!?
Could have been very incriminating, I wonder if she could copy the code?
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
- Some mornings I wake up cranky.....but usually I just let him sleep in -
Morse Code - too much of it will make you "dotty"
I bet the typos in my previous post had been:
1. Unimportant
2. Annoying
3. Un-noticed
4. Filled in or corrected automatically because you "understood" the context
5. Re-Read,. verified as having errors then...
Meaninful extrapolation of data is something computers can't do without the storage capacity available to consider all the required possibilities. Regular or extended expressions can generate oodles of patterns but without a reference to determine what is useful in context is wastefull. Time taken to do this does not really matter unless you can't wait for an answer - either right, wrong or unknown. Impatience is why most development efforts followed the "clock speed" path - doesn't mean diddly compared to parallel processing. Yes, I have an X MB of cache multiplied by a Y SMP environment - guess what, an external device has just updated your database; all your fancy caches are now stale - go refresh!
How many words begin with ""W", end with a period and contain between 2 and 9 letters? Does it matter - why or why not?
BTW: I'm not suggesting investing any money in anything; I'm just being open-minded that someone may have found a "better way" of doing things we thought we were already doing; not to imply that what we already do is the best way to do it - that's all.
Cheers,
Harry E.
First, they released a video extolling "Dr. Jim"'s theory of how neurons work, which is simply wrong. Neurons don't work the way he says, even accounting for the fact that he might have made up his own lingo and be using words differently than other researchers. Neurons are not "on or off." There is no reasonable interpretation of neural function that can be reconciled with the ideas in that video; it is not avant-garde, it is simply wrong, and if the claim is that it will work because it is mirroring the example of biology, well that's wrong too because it's not mirroring the example of biology.
Then in this more recent example we get a description of how speech will be recognized, which starts by throwing away most of the information which people have known is there since roughly 1940. Again, this is not "new" or "better," it is simply wrong. There is no way to interpret that blog post even assuming very liberal definitions for things like "carrier" that is at all consistent with what is known about the structure of speech. While one might say MIT are inventing a new wheel because they think the real scientists are doing it wrong, it's kind of hard to justify that when we know for a fact that the human ear also does spectral separation.
I have met some of the people who are doing serious work on AI. There is FAR more to it than pattern recognition, and it is an insult to a lot of people who have literally dedicated their lives to the task to suggest they are too stupid to figure out that stale caches are the problem. (In fact, your statement of that suggests that you don't have a clue how caches work; there's a lot of fancy math involved in making them transparent to the very kind of thing you describe.)
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Post Edited (Leon) : 8/23/2009 2:32:11 PM GMT
As for quantum processing, "far-fetched" is about the kindest thing I can say about that sort of thing. I do think that plain old thermal noise has a lot to do with it, possibly in a way that is necessary rather than just murking up otherwise predictable results. The the quantum silliness harks back to the notion that the brain is so powerful and unpredictable that it can't possibly be something simple enough to emulate with a mere computer. I do believe that's poppycock; there are ten billion or so neurons in a human brain, each with thousands of interconnections. Even if an individual neuron is a fairly straightforward thing to emulate (which I think it ultimately will be) that's a lot of complexity compared to the computers we've built so far.
I don't see the relevance of thermal noise to communication between neurons, the effect would be negligible.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Post Edited (Leon) : 8/23/2009 3:09:57 PM GMT
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
With apologies, the references were not intended to insult anyone but appear to have done so but I don't see any implication of anyone being "stupid".
References were purely intended to illustrate the _time_ needed for a CPU to flush cache (1st, 2nd or Nth level) and access (any) external [noparse][[/noparse]bus or I/O controller based] memory, multiplied by the number of CPU's on-line in an SMP (Symmetric Multi-Processor) environment to maintain [noparse][[/noparse]cache] coherency; should this be necessary.
WRT "pattern-matching"; this was perhaps an over-simplified example. To expand; the reference was WRT distilling of data and patterns representing 'sentences/commands or whatever' that can then be decoded/replayed/looked-up using appropriate methods.
If one was to encode the "simplified" aural digitisation of "Go to the refrigerator and get a cold drink" into a sequence of a minimal set of bits (with an attached EDC/CRC - or whatever) - it would be, could be or definitely should be unique - if not the encoding mechanism should compare what it encoded with what it "knows" and if it reveals an existing pattern for a (perhaps) different "thing" then it should seek clarification of what constitutes differences. Better that than "go drive the car into the river" because it "thought" it understood.
Finally, the simplistic cache reference was allowing an "external 'device'" to update {memory} - a common resource in this context which of course would have to use either a DMA mechanism; requiring some form of "bus mastership" ability, or, another form of access; controlled by some form of "handler". In either case, any CPU (SMP or other) with a cache mechanism would have to be advised their references to data for addresses affected by the "DMA" cycle would now be invalid [noparse][[/noparse]if any].
This could include a Propeller if they were processing real (or near) time data related to environment and or activity - physical or otherwsie; using data obtained from a "common" source.
Cheers,
HarryE.
Harry -- Again, you vastly oversimplify something that has been studied to death.· The time needed for a modern CPU to flush cache is generally negligible because (1) the cache is in small bits, the whole thing isn't flushed and refreshed at once, (2) memory hardware is optimized for block read and write, and most important (3) in a properly designed system cache refresh is usually performed via read-ahead and write-behind while other bits of the cache are in use by the CPU.· This has all been standard in PC's for at least 15 years.
As for aural digitization, we have known since the 1940's both that the "encoding mechanism" for phonemes is a time-variant spectrum, and when you throw away or never extract the spectral information, as MIT proposes, you are not "encoding" anything.· Speech recognition is not actually a total mystery, and we know how the basics work; MIT proposes to ignore all this and take a path which we frankly know doesn't work.· You cannot compare information you have thrown away to anything.
We have also had some time to learn to build massively parallel computers: http://en.wikipedia.org/wiki/Connection_Machine· An interesting fact about the early Connection Machines was that their individual processors were too simple to have caches.· Now considering what these machines are used for and the progress made in the last 25 years, do you really think the people spending millions of dollars to build these things are so stupid that they are going to be tripped up by the equivalent of "OH NOES I GOTZ 2 FLUSH MAH CASH?"
It comes from learning (and needing or wanting) the difference. How much trivia can you recall - given something you find interesting (ie: important or useful) versus "Facts and Figures of 800 BC" you were forced to sit through by your Library Teacher when you were 9?
When you started learning, primitives were set; as were doctrines, diciplines and right and wrong. Being able to determine right and wrong goes a long way; Doing something that is "wrong", fun and worth the subsequent "punishment" is a good risk and worth more if you don't get caught and suffer the punishment - right?
All references are based on the "dance" (choreography) of the actions of determining all avalable stored and real-time data that is used in choosing what to do (be it right or wrong) [noparse][[/noparse]this of course forms the basis of an AI machine - it is not allowed to do wrong because Humans say so].
However without the continual storing of "experiences" [noparse][[/noparse]input data] and filing for "use{ful,less}" recall later, you (not personally but any-critter) are less capable or more likely to make a "bad call" - especially as the details surrounding the "call" get more complex. This is sometimes defined as "experience" - as they say "it comes with age" - why? I've seen people make the same mistakes multiple times for similar reasons - how come? We are supposed to learn from mistakes - right?
Walk down the street on a nice Summer day - a smell triggers a memory - long since forgotten (buried or whatever) - why yes but it's still there in _every detail_ - why and how?
Cheers.
HarryE.
Of course
It's the simple things that'll get ya - every time. Some people have become so impressed by their own "intelligence" they sometimes forget their place. One of the Forum participants has a nice signature line - simplicity.
What reasons would there be here for not using simplistic references?
Cheers,
HarryE.
Thermal noise has a negligible effect on brain neurons. If it did have an effect they couldn't work properly. Alopex has nothing to do with real neurons.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Post Edited (Leon) : 8/23/2009 5:15:19 PM GMT
-Phil
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Alopex was based on real neurons, specifically the architecture of the Thalamus. And it isn't the only algorithm in this family that requires a random input. Your assertion that noise has nothing to do with how neurons work, right after you asserted that we don't know how neurons work, strikes me as more than a little strange.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Post Edited (Leon) : 8/23/2009 8:26:47 PM GMT
And having lived long enough to experience what age starts to do to one's brain, I would advise that cell death and noise are probably major parts of the process whether they're necessary or not.
Is there any evidence that neurons actually use noise in their operation?
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Post Edited (Leon) : 8/23/2009 8:56:17 PM GMT
Then, if I keep thinking, the rest of the mechanism keeps it going just like an automatic watch.
··
· "I don't think it's the high fever that causes the delusions, it's more likely to be caused by the the toxins that are causing the fever."
LOL
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
BH
Post Edited (BADHABIT) : 8/24/2009 10:28:35 AM GMT
This sounds like some kind of chemotaxis to me en.wikipedia.org/wiki/Chemotaxis all be it if the neurons are not actually moving, they are reaching out in some direction.
Which poses the question, how does a neuron know in which way to grow its synapses? What is the chemical gradient or chemotactic signaling they are following? And how did that come to be there?
When I here people suggesting that a neuron is "a simple thing" that can be modeled in a computer algorithm and many instances of it combined into a neural net which will then model a brain, I never here any mention of this chemotactic growth phenomena which seems to already be quite a complicated thing.
Looks to me like a single neuron has a lot of complicated, even "smart" behavior that is overlooked. It's at least as smart as that bacteria swimming around searching for food and avoiding harmful environments.
Still, what do I know.
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
For me, the past is not over yet.
Leon
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle
Post Edited (Leon) : 8/24/2009 11:17:34 AM GMT
When you said that the morse tappings were "used to drive his wife crazy" did you mean that he did it innocently and it upset her, or, it WAS USED to drive her crazy. The last Mrs Seckshund went that way all on her own accord, I think,·as I never learnt morse (only a G8).
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Style and grace : Nil point