News Marvin Minsky, one of the pioneers of artificial intelligence, has died of a cerebral

[Notebook]

http://www.bbc.co.uk/news/technology-35409119

The mathematician and computer scientist was one of the world's foremost AI experts.

As a student, he built one of the first neural-network learning machines, using vacuum tubes.

He went on to cofound the Massachusetts Institute of Technology's Artificial Intelligence Lab, in 1959, with John McCarthy.

Prof Minsky's ideas and influence were wide-ranging - from computational linguistics, mathematics and robotics - but underpinning it all was a desire, in his own words, "to impart to machines the human capacity for commonsense reasoning".

 

[Ravenous]

I can't find the reference, but I think he was the first to coin the phrase that brains are "Computers made of meat."

His point being that brains can (at least in principle) be fully replicated by a conventional computer. Still a controversial idea (don't ask me for an opinion, I simply don't know!)

 

[Col_Klonk]

His point being that brains can (at least in principle) be fully replicated by a conventional computer. Still a controversial idea (don't ask me for an opinion, I simply don't know!)

It's just that we need the parallel processing capability of the brain put into silicon (approx 100 billion neurons at adulthood). These parallel cpu's haven't really taken off as they have been a bit slow, and the same functionality could be implemented in software on an Intel cpu, and it would be faster.

But later Parallel architecture cpu's began to arrive....
The Inmos Transputers were the first parallel cpus, now Xmos cpu's are the next attempt. There is a neuron specific CPU.. I forget the manufacturer and name.. (it was also slow at the time thus not too popular).

Generally one can (with a bit of work :thumbup implement the same idea into FPGA's and the logic would be definitely create a faster neuron logic.
There's has been development in asynchronous FPGA's (FPGA's are usually clocked like our desktop cpu's) which would suite this field.

The problem is that you need a LOT of CPU's or massed logic.
the last time I looked, I think we've only got as far as the brain of a flea (or fly).

 

[Urwumpe]

Well, the human brain isn't essentially effective, but pretty flexible and relatively fast. I doubt that you really need to replicate a silicon brain to get intelligence, what you need is just the way it works. Which is not massively parallel, just many basic tasks are done by special processors, the rest more looks like a map-reduce algorithm.

 

[Col_Klonk]

Essentially yes..

As we know the brain get's it's power from massively parallel operations, but this also includes a memory system using neuron feedback.

Current computers have these basic systems.. and with the speed of cpu's seemingly faster than the neuron.. we must just create the right 'algorithm' to simulate massively parallel in smaller parallel systems running in serial mode.
Computer memory size and access speed might be a problem though.

 

[Urwumpe]

Essentially yes..

As we know the brain get's it's power from massively parallel operations, but this also includes a memory system using neuron feedback.

Current computers have these basic systems.. and with the speed of cpu's seemingly faster than the neuron.. we must just create the right 'algorithm' to simulate massively parallel in smaller parallel systems running in serial mode.
Computer memory size and access speed might be a problem though.

Not sure. I think the human memory capacity is over estimated, we just have a good and very error-prone compression algorithm. For example, we don't remember video. We remember models and events. Aside of the things we really focus on, our memory has to fill gaps with things we already knew before.

The way how we turn what we see into models is the mystery. It is really hard to replicate, though neural networks are getting close to it.

 

[Artlav]

The computational power is likely overestimated as well.
Contemplate this - a modern CPU have billions of transistors in it.
How would you simulate it on a different CPU?

Simulate every transistor? It would run at an instruction per second, at best.
Simulate the logical elements these transistors form? You might get a 1000 instructions per second.
Simulate the instruction set these logic gates form? That would be millions of instructions per second.
Rewrite the software that is running on it to the target CPU's instruction set? That would be billions of instructions per second, or even more.

Just this simple task vary by a dozen of orders of magnitude depending on how you approach it.

So my guess is that if we really knew all the details of the algorithms behind the intelligence, we might be able to run a human realtime on something as small as a gaming-grade desktop computer, give or take a few ASICs.

 

[Urwumpe]

So my guess is that if we really knew all the details of the algorithms behind the intelligence, we might be able to run a human realtime on something as small as a gaming-grade desktop computer, give or take a few ASICs.

Exactly. We must learn how to design a brain for task. Not how to copy one without understanding it.

 

[Col_Klonk]

That's more or less where they started this AI stuff.

Another thought is that one could use the Graphics processor units for massively parallel processing. Nvidia seems to be the ideal option for this compared to AMDs stuff.

 

[Urwumpe]

Another thought is that one could use the Graphics processor units for massively parallel processing. Nvidia seems to be the ideal option for this compared to AMDs stuff.

Actually wrong. :lol:

The AMD architecture is generally more capable in that field. AMD has based their architecture around classic SIMD calculations, while NVidias Maxwell is optimized for executing chained CUDA and user-defined kernels.

Of course, both manufacturers claim to be the best, regardless of the software that runs on them. :lol:

http://devblogs.nvidia.com/parallelforall/maxwell-most-advanced-cuda-gpu-ever-made/

The big disadvantage of AMD though: Their approach requires careful optimization of the compiler, to make sure that their SIMD units are effectively used. The NVidia approach does not require that much intelligence on that end, but is more limited in the calculations.

Essentially, it is about the calculations that you want to run, to tell what will be more effective.

You can find it in the TOP500 list - while NVidia is leading in the top ten, its also the systems with the largest number of cores and the most expensive systems. But behind those state-funded monsters, you can see that AMD gets a better GFLOPS per core.

 

[Col_Klonk]

Nvidia dev kits are much better, and Nvidia seems to offer more info.. when looking at AMD's stuff ect (both regularly for the past 2 years).. I chose Nvidia.
This wasn't bias choice either

 

[Andy44]

So if a not-so-complex machine could do it, could you "reprogram" the brain of a chimp or a dolphin to "run human"?

I need to write that down...

 

[Linguofreak]

You can find it in the TOP500 list - while NVidia is leading in the top ten, its also the systems with the largest number of cores and the most expensive systems. But behind those state-funded monsters, you can see that AMD gets a better GFLOPS per core.

Of course, part of NVidia's dominance there might be because supercomputers tend to run Linux and NVidia generally does a better job with driver stability on Linux than AMD. Of course, I'm sure AMD pays a lot more attention to Linux driver quality on supercomputers than it does on the desktop, but I do wonder to what degree that is a factor.

 

[Urwumpe]

Of course, part of NVidia's dominance there might be because supercomputers tend to run Linux and NVidia generally does a better job with driver stability on Linux than AMD. Of course, I'm sure AMD pays a lot more attention to Linux driver quality on supercomputers than it does on the desktop, but I do wonder to what degree that is a factor.

Well, you also have to differ between graphics and computing drivers there, graphics drivers on Linux is a story with many misunderstandings, while the scientific department works pretty fine for both manufacturers.

Still I think that the difference primarily is, that NVidia got on the supercomputer stage first. AMD did not have something competitive until some years ago, while NVidia essentially defined the market. When most TOP10 systems had been defined and procured, AMD did not really exist for such applications, except by small cluster projects in universities.