• Quick note - the problem with Youtube videos not embedding on the forum appears to have been fixed, thanks to ZiprHead. If you do still see problems let me know.

Super Artificial Intelligence, a naive approach

Reality Check said:
The actual calculation shows that it is impossible to achieve your claim by 2020 of roughly 10^16 to 10^18 synaptic operations per second.
The OP links to a 2014 IBM SyNAPSE chip with "only" 256 million synapses.
A claim of 10^14 synaptic operations a second gets to the lower limit of the range that you gave in a little over 8 years (doubling every 2 years), i.e. at least 2025. But your source for this is not an currently working chip - it is a simulation on a massively parallel supercomputer.

The rest of the post may need a Duh! because it is just about inevitable that we will have "a brain in a box" sometime - maybe within the next few decades.
Click to expand...

(A)
Don't be afraid of the word "simulation".

Simulations may yield proper outcomes.

For example, alpha go, the planet's prominent ai, and the planet's initial approximation of general artificial intelligence, used SIMULATIONS of scenarios to acquire its experience.

The use of SIMULATIONS did not stop it from destroying Lee Sedol, the planet's human go champion (before alpha go that is)



(B)
IT IS NOT 10^14 SYNAPTIC OPERATIONS PER SECOND.
People in this forum continue to repeat that error.

It is 10^14 SYNAPSES.

One is SYNAPTIC OPS PER SECOND, and the other is the SYNAPSES themselves.


(C)

This is why the original post began with 10^16 sops.

I could have began with 10^15 synapses, instead of 10^16 sops.

The 10^15 synapses is rough for some value x 10^15, or rough for 10^16 sops.

People here even till now, still don't get that 10^15 SYNAPSES can be ROUGH for 10^16 SOPS. (They still ignore the units )
 
Last edited:
Reality Check said:
Do not be afraid of the word chip! Or working!
You need to read your OP :eye-poppi!
You wrote "(iv) ... 10^14" and link to an irrelevant IBM chip so it is 10^14 computer operations a second.
Click to expand...

Once more, the simulation entailed 10^14 synapses.
So, the next iteration could entail 10^16 synapses, regardless of the substrate.

The point is the calculation could take place on the cardinality of synapses simulated.

So your initial point of 10^14 SYNAPTIC OPERATIONS per SECOND, did not exist in the document analysed.

You can choose to ignore your error or not.
 
Last edited:
Last edited:
BTW, @RealityCheck, you may be the first being qualified to converse about the topic at hand. What do you make of the following?

Note: you can answer if you have experience in regression work. (I know you mentioned you have experience on manifolds)

ProgrammingGodJordan said:
Here is slightly better description, in loose machine learning terms:

  • Points maintain homeomorphisms, such that for any point p under a transition T on some transformation/translation (pertinently continuous, inverse function) t, p0 (p before T) is a bijective inverse for p1 (p after T); on t.

  • Following the above, topologies maintain homeomorphisms, for any collection of points W (eg a matrix of weights), under some transition T on some transformation/translation sequence (pertinently continuous, inverse functions) s, W0(W before T) is a bijective inverse for W1(W after T); on s, where for any representation of W, determinants are non-zero.


  • Now, topological homeomorphisms maintain, until linear separation/de-tangling, if and only if neural network dimension is sufficient (3 hidden units at minimum, for 2 dimensional W)

    Otherwise, after maintaining homeomorphism at some point, while having insufficient dimension, or insufficient neuron firing per data unit, in non-ambient isotopic topologies that satisfy NOTE(ii) W shall eventually yield zero determinant, thus avoiding linear separation/de-tangling. At zero determinant, unique solutions for scalar multiplications dissolve, when the matrix becomes non-continuous, or non-invertible.

NOTE(i): The state of being "ENTANGLED" is the point before which some de-tangleable classes are de-tangled/made linearly separable.

NOTE(ii): Unique solutions in matrices are outcomes that resemble DATA SETS; for homeomorphisms (topologies: where zero-determinant continuous invertible transformations/translations engender OR ambient isotopies: where positive/nonsingular determinants, nueron permutations, and 1 hidden unit minimum occurs, i.e for 1-dimensional manifold, 4 dimensions are required in network)


[qimg]http://i.imgur.com/oIOuGxD.png[/qimg]
Click to expand...
 
Reality Check said:
The error was in your OP as you know: No 10^14 in that link :jaw-dropp

I am the one who fixed your mangled quote about the irrelevant simulation: 10^14 synapses for current machine level wrong because no current machine has 10^14 synapses.
Click to expand...

(A)
I still observe that IBM simulated 10^14 synapses in 2012.

Moore's law is not limited to non simulated units.


(B)
Anyway, your statement there is no machine with 10^14 synapses is invalid.

Being simulated doesnt magically erase the 10^14 synapses.

(C)
Your "correction" merely enforced your ignorance.
Simulations do not imply inexistence.
 
Last edited:
Reality Check said:
Random highlighting does not make a better description of anything. What might be a cut and paste from a textbook is a waste of space - link to the source.
Click to expand...

Not random.
Also, cut/paste is appropriate; for I am the author, if you look a bit.

Reality Check said:
ETA: looks more like a cut an paste from here which hints of out of context, mathematic word salad from an amateur.
Click to expand...

Nitpick: Why the redundancy? (as mentioned in the link you referenced/described, the description is amateur)

If possible, could you provide us with the correct representation?

Otherise, if you can't, how do you know it is "mathematic word salad"?
 
Last edited:
Reality Check said:
I added some text to the posts you replied to but I will emphasize this:
10^14 synapses for current machine level wrong because no current machine has 10^14 synapses.

Even with this imaginary 10^14 synapse machine, you fail to get to 2020.
10^14 * 10 = 10^15 synaptic operations per second.
Double this to get to 2019: 2 * 10^15 synaptic operations per second.
Double this to get to 2021: 4 * 10^15 synaptic operations per second.
That is less than half of the lower limit of your claim. So the claim is debunked even if we ignore the reasonable interpretation of "roughly" + a range as meaning "somewhere in that range" :eek:!

Even worse, the mostly likely value in a range of values in the middle of the range, thus:
It is standard in science and common in real life that a calculation on a range of values does not use the extremes or values outside of that range
Click to expand...

FORMULAE : HBS = CMS * 2^n

HBS=human_brain_speed (where size corresponds with speed, but I use the size)
CMS=current_machine_speed (where size corresponds with speed, but I use the size)
n = YEARS_TILL_BRAIN_CHIP/rate
RATE = 2


So,
HBS = 2*10^15
CMS = 6.4*10^14

So,
(2*10^15) = (6.4*10^14) * 2^(YEARS_TILL_BRAIN_CHIP/2)

(2*10^15)/(6.4*10^14) = ( (6.4*10^14) * 2^(YEARS_TILL_BRAIN_CHIP/2) )/(6.4*10^14)

(2*10^15)/(6.4*10^14) = 2^(YEARS_TILL_BRAIN_CHIP/2)

3.125 = root (2^YEARS_TILL_BRAIN_CHIP)

YEARS_TILL_BRAIN_CHIP = 3.28




So, we roughly have 3.28 + CURRENT_YEAR, which gives us roughly 2020.
 
Reality Check said:
It is standard in science and common in real life that a calculation on a range of values does not use the extremes or values outside of that range. The reasonable value to use is a value in the middle of the range. That is usually the average or median value.
Click to expand...



This is why the original post began with 10^16 sops.

I could have began with 10^15 synapses, instead of 10^16 sops.

The 10^15 synapses is rough for some value x 10^15, or rough for 10^16 sops.

People here even till now, still don't get that 10^15 SYNAPSES can be ROUGH for 10^16 SOPS. (They still ignore the units )
 
Reality Check said:
Wrong: Moore's law is the observation that the number of transistors in a dense integrated circuit doubles approximately every two years.
You do know what an "integrated circuit " is?
Click to expand...

You're talking to the guy who calls himself a god because humans can make computer models of the Universe. I don't think he distinguishes between a simulation and the actual thing.

Also, he still hasn't given a proper citation for his assumption of 10 SUPS in the human brain. Only a link to an article that admits it's a guess and obliquely references a 1985 book.
I don't have that book, but I have found it on Google Books. The search function is very limited, unfortunately, And the only reference to synaptic speed I've been able to find there is a reference to motor neurons sending between 5 and 100 signals per second.
This could very well be a failing on my part, but as long as OP has nothing more than a 1989 article by a computer scientist, I'm not going to take his word on matters of neurology.
 
ProgrammingGodJordan said:
People here even till now, still don't get that 10^15 SYNAPSES can be ROUGH for 10^16 SOPS. (They still ignore the units )
Click to expand...

We've asked it time and again.

Apart from a single reference in a 30 year old article (where it was a "guess"), where is your evidence that 1015 synapses are equivalent to 1016 synaptic operations per second ?

Also:

Why have you glossed over the fact that the IBM claim for 1014 synapses in a simulation which was running "only" 1542 times slower than realtime ?

Rather than it being a factor of 3 that you're trying to make up, it's a factor in the thousands, tens of thousands or hundreds of thousands that you'll need to make up....

.... and of course there's significant doubt that Moore's Law will still apply as physical limits are reached.

Oh, and even if we somehow get to the "right" number of synaptic operations per second, it assumes that the "software" running is of comparable efficiency to the human brain.
 
What does 10^15 or 10^17 synapses get you? A lot of synapses.

A novel has roughly 10^6 alphanumeric characters. But if you string together 10^6 alphanumeric characters, you don't necessarily have a novel.

An Intel i7 processor has roughly 10^9 transistors. But if you connect together 10^9 transistors (say, three semi truckloads of 2N2907s), you don't necessarily have an i7 processor.

An elephant has roughly 10^28 amino acid molecules. But if you pile up 10^28 amino acid molecules, you don't necessarily have an elephant.

There's the part about connecting the things together in the right configurations to consider. I assume you wish for the ten-to-the-whatever synapses to produce the outward effects of human intelligence instead of, for instance, the outward effects of a human coma. How long does that part take?
 
Myriad said:
What does 10^15 or 10^17 synapses get you? A lot of synapses.

A novel has roughly 10^6 alphanumeric characters. But if you string together 10^6 alphanumeric characters, you don't necessarily have a novel.

An Intel i7 processor has roughly 10^9 transistors. But if you connect together 10^9 transistors (say, three semi truckloads of 2N2907s), you don't necessarily have an i7 processor.

An elephant has roughly 10^28 amino acid molecules. But if you pile up 10^28 amino acid molecules, you don't necessarily have an elephant.

There's the part about connecting the things together in the right configurations to consider. I assume you wish for the ten-to-the-whatever synapses to produce the outward effects of human intelligence instead of, for instance, the outward effects of a human coma. How long does that part take?
Click to expand...

There is indication that current configurations are non-trivial, but mind you, as mentioned in the original post, merely human level brain power (i.e. the cardinality of cycles per moment) shall probably be roughly achieved (i.e. artificial human level intelligence in all degrees was not mentioned for 2020)

It is observable, that today, that these cognitive models either match human performance in cognitive tasks, or exceed. (Eg alpha go, disease diagnosis neural nets, etc)

It would also be silly to ignore that as time/parallelism enhances, these cognitive models do more and more cognitive tasks, well, and have already caused job displacement, and shall probably cause more.
 
Last edited:
The human brain computes on average at 10^17 synaptic operations per second

ProgrammingGodJordan said:
This is why the original post began with 10^16 sops.
Click to expand...
Yes - the OP has "(iii) The human brain computes roughly 10^16 to 10^18 synaptic operations per second" - needs a Duh :D!
The rational value to use is something in the middle, e.g. 10^17 synaptic operations per second according to your numbers.

Taking an OP as correct usually leads to a bit of a derail. In tis case case I took your "(iv) Mankind has already created brain based models that achieve 10^14 of the above total in (iii)." to mean that you had a reference to 10^14 synaptic operations per second as in (iii).

Actually you did not.
The OP has a wrong link to a 256 million synapse IBM chip. No 10^14 sop there. Later you correct the link to an IBM Research Report from 2012 (PDF)
Since the final submission of our work on the Compass scalable simulator for the IBM True North Cognitive Computing architecture [1], we have simulated an unprecedented 2.084 billion neurosynaptic cores containing 53 x 1010 neurons, and 1.37 x 1014 synapses, running at only 1542x slower than real time. ...Shepherd [2] estimates the number of synapses in the human brain as 0.6 x 1014, and Koth [3] estimates the number of synapses in the human brain as 2.4 x 1014.
Click to expand...
That report is about a computer simulation containing 10^14 synapses. That is not 10^14 synaptic operations per second. There is no statement of any "operations per second" in that report.

Therefore the simulation of 10^14 synapses cannot be compared to any speed of any computations.
 
Last edited:
Reality Check said:
Yes - the OP has "(iii) The human brain computes roughly 10^16 to 10^18 synaptic operations per second" - needs a Duh :D!
The rational value to use is something in the middle, e.g. 10^17 synaptic operations per second according to your numbers.

Taking an OP as correct usually leads to a bit of a derail. In tis case case I took your "(iv) Mankind has already created brain based models that achieve 10^14 of the above total in (iii)." to mean that you had a reference to 10^14 synaptic operations per second as in (iii).

Actually you did not.
The OP has a wrong link to a 256 million synapse IBM chip. No 10^14 sop there. Later you correct the link to an IBM Research Report from 2012 (PDF)

That report is about a computer simulation containing 10^14 synapses. That is not 10^14 synaptic operations per second. There is no statement of any "operations per second" in that report.

Therefore the simulation of 10^14 synapses cannot be compared to any speed of any computations.
Click to expand...

(1)
At 10 impulses per second, and 10^15 synapses (based on older estimations, and some newer estimations for the child brain), we get 10^16 sops, so OP is okay, at 10^16 to 10^18 synaptic operations per second.

Where did you get 10^17 from?


(2)
OP did not express 10^14 sops, but 10^14 of the above range in (1). I later mentioned many many times, that 10^14 synapses of the above range, rather than sops were used.



(3)
From (2), ibm contained 10^14 synapses, as mentioned extensively throughout this thread, and mentioned in prior responses to you.



(4)
I linked to IBM's website, with the relevant url.



(5)
How silly.

Why can't synapses (simulated or not), be compared to sops, when sops are reducible to synapses?

An above average toddler could make the above comparison. (although this is performable by established computational neuroscientists)




(6)
Challenge yourself, and ignore the trivial 2020 estimation (such is not this thread's core topic, as I revealed many times prior). Let us discuss manifolds/super-manifolds, as it relates to neural modeling.
 
Last edited:
You must log in or register to reply here.

Back
Top Bottom