Re: Machine takeover?

Tim Tyler <tim@xxxxxxx> wrote:
Curt Welch wrote:

A brain has 100 billion neurons. A big server with 4 Gb of memory has
4 billion transistors in the memory chips alone before you add in all
the circuits of the rest of the system. And those transistors operate
at information processing speeds which are a few orders of magnitude
faster than the big slow, information processing power of the complex
living cells called neurons. 4 billion processing elements operating at
speeds 100 times faster puts in you in the same ball park as 100
billion slow neurons.

A brain has (100-500 trillion) synapses - which are probably more closely
comparable to transistors than neurons are. That's quite a few.

Yes, that's an important question to be resolved. Why is the brain using
so many activate signal processors while processing so little total
information? And most important to AI, are their any alternative
implementations that would allow us to build the equivalent signal
processing function, using far less signal processors?

The response time requirements of real time behaviors (like recognizing
prey and predator) limit the length of the processing chain because of how
slow neurons and nerve fibers are at moving data. It has created a system
that can produce complex answers by passing large amounts of data though a
parallel chain that only has a depth of about 5 neurons (based on how was
humans and animals can produce a complex response). In order to survive,
evolution was forced to add more complex not by increasing the sequential
length of the signal processing chain, but by making it much wider - making
it highly parallel. I suspect however that also makes it highly wasteful
of synapses. That it had to add lots of cross connections, to make up for
the fact it couldn't add any more depth to the signal path.

I strongly suspect, that since transistors and wires are many orders of
magnitude faster than neurons, that we can use an architecture that uses
more depth, and far less cross connections, and come out with a huge
reduction on the number of transistors (compared to how many synapses the
brain had to use).

I say what because a prime aspect of what the brain has to do, is act as a
signal switching system. It has to switch information from millions of
different inputs to thousands of different outputs very quickly. It does
it now in a network of very limited effective depth, by using a huge number
of cross connection paths.

If you have 1000 inputs, and 1000 outputs, and must have the power to
switch any input to any output in one layer, you have to have 1000x10000
(100,000) cross connections (every input to every output). But if you can
do it in multiple steps, you can build a switching network with far cross
connections which I believe approaches N log N connections for the most
optimal configuration instead of N^2. So 3000 instead of 100,000 in this
example.

I suspect the architecture of the brain was forced to use a configuration
that was closer to the N^2 worse case of a single level, then the best case
N log N result if you could use the log N number of layers instead of 1
layer (or 5 which it seems is closer to what the brain was able to use in
solving the switching problem).

That effect might allow us to produce a solution with the same effective
power as the brain, but with many orders of magnitude, less switching
components.

But of course, this is just more speculation about what might be. We might
need 500 trillion transistors as well. That would make it large and
expensive by today's standards, but still cheap enough that we could afford
to build one today if we knew what to build.

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Relevant Pages

  • Re: Consciousness: whats the problem?
    ... I take it to be a property of the functioning brain. ... If the behaviors can be accounted for without appeal ... Each of the 10^11 neurons has ... You have to add in the fact that the bandwidth of wires and a transistors ...
    (comp.ai.philosophy)
  • Re: What is the state of Robotics Currently.
    ... estimate the amount of computing power to duplicate human intelligence ... We don't need to simulate transistors to create a computer, ... with neurons switching at 1000 times per second and 100 ... Large parts of the brain is filled with interconnects (white ...
    (comp.robotics.misc)
  • Re: Strong AI Thesis (No Chinese room, I promise)
    ... My objection was that you are using the word "existence" differently from ... What also exists is millions of other brains that have hardware similar to ... At the next level of abstraction, what exists is hardware in your brain, ... limiting your vocabulary to only transistors. ...
    (comp.ai.philosophy)
  • Re: Superintelligence: Singularity and the Approach to Superhuman
    ... Microprocessors have raw computational power. ... When we dig down into the guts of a high-speed modern microprocessor, what we find are a bunch of very simple two-state devices -- tiny little transistors that are either 'on' or 'off'. ... Now let's think about the brain. ... Add to that that one neuron doesn't just connect to one neuron, it connects to many on both ends. ...
    (talk.origins)
  • Re: Sony KV-35S45 Power Supply
    ... course blew the brand new switching IC. ... normal switching transistors, similar to those used in other Sony ... diode from the secondary side. ... This is a difficult circuit to troubleshoot ...
    (sci.electronics.repair)