Scientists to Build 'Brain Box'

lee1 writes "Researchers at the University of Manchester are constructing a 'brain box' using large numbers of microprocessors to model the way networks of neurons interact. They hope to learn how to engineer fail-safe electronics. Professor Steve Furber, of the university school of computer science, hopes that biology will teach them how to build computer systems. He said: 'Our brains keep working despite frequent failures of their component neurons, and this "fault-tolerant" characteristic is of great interest to engineers who wish to make computers more reliable. [...] Our aim is to use the computer to understand better how the brain works [...] and to see if biology can help us see how to build computer systems that continue functioning despite component failures.'"

Comments

Downside of biological computing

[QuantumFTL]

While I was an intern at the Jet Propulsion Laboratory, back when I was an undergraduate, I was very gung-ho about biologically inspired computing - I implemented an automatic flowchart positioning system using a genetic algorithm that would "evolve" a correct solution to the problem. While this certainly worked to some extent, the instability and sheer unpredictable nature of using such a stochastic algorithm made it impossible to use in a mission-critical setting. Many biologically inspired algorithms solve problems through methods that cannot be proven correct (unlike, say, the mathematics circuitry in a CPU), but merely empirically observed to "do a good job."

One of the main drawbacks of human engineering is the need for certainty, which often prohibits the use of many high-efficiency stochastic algorithms (especially for things like mesh communication) in conservative industries, like the US defense industry. This is also a significant problem in other areas, however, and many biologically inspired algorithms have properties that we cannot, so far, completely explain - they are treated like "black boxes" with many unknowns for engineering purposes.

I think that in certain circles, the tremendous success that is evolution on this planet has overshadowed its enherent weaknesses - that it is a greedy, local optimizer which cannot reach a large amount of the possible biological search space due to being stuck in local optima, and the added constraint that everything must be constructed out of self-replicating units (these two factors are why something useful, like, say, a Colt 45, will never emerge without the pre-existence of an intelligence). Biological examples are fascinating and often practical, but the biological approach is almost always "brute force" and/or "sub-optimal but still alive."

I think biologically-inspired algorithms will continue to gain prominence, but in my estimation, it is likely that there will be harsh limits imposed on how far guarantees of performance from empirical tests and symbolic analysis will actually hold.

Re:Testing for fault tolerance

[CroDragn]

This has been done before, introducing a random element into the neural net. If done correctly, this can result in "creativity". Here is one link about it, seen it many other places too, so google for more.

some amusing calculations

[llZENll]

well the article is so short its not possible to comment on their implementation. so here are some calculations i did to amuse myself.

number of neurons in the brain: 100 billion
http://hypertextbook.com/facts/2002/AniciaNdabahal iye2.shtml

transistor count per CPU: ~300 million
http://www.anandtech.com/cpuchipsets/showdoc.aspx? i=2795

average synaptic connections per neuron: 7000
http://en.wikipedia.org/wiki/Neuron

total number of synapses: 100 to 500 trillion

since a 'calculation' for one artificial neuron mostly involves a summation of weights, we can view one total step as 2 X the number of synapses we wish to analyze. or 200 - 1000 trillion calculations for one step. by step i mean summing all inputs and pushing the result to an output for each neuron.
http://en.wikipedia.org/wiki/Artificial_neuron

fastest computer in the world FLOPs: 280 trillion
http://en.wikipedia.org/wiki/Blue_Gene

pentium 4 FLOPs: 40 GFLOP

using the fastest computer in the world 1 step would only take around 1 - 5 seconds, not counting storing all of that information.
http://en.wikipedia.org/wiki/Blue_Gene

so how fast do we think? well i couldn't find anything on this so lets get a quick estimate. the average neuron is .1m in length .1 / c = 3.3x10^-10 or 333 picoseconds. now lets add in some delay for the chemicals in the neurons to do their thing, this is probably much slower than the electrical impulse, so lets say 3.3 nanoseconds.

so assuming our computers could network instantly, and store the data used instantly, we would need 3-15 trillion Blue Gene supercomputers to simulate the human brain in real time. or if we are using pentium 4s we would only need 21-105 trillion pentium 4s.

man thats a lot of cpus.

number of computers in the world: ~300 million
http://www.aneki.com/computers.html
guess at average FLOPs per computer: 40 GFLOPs
total FLOPs of worlds personal computers: 1.2 PFLOPs
time to calculate one brain step if all computers in the world were networked: .2 - .8 seconds

using moores law, when will a single computer be fast enough to simulate the human brain in real time?
200-1000 trillion calculations per step = ~600 trillion every 3.3ns = 181x10^18 or 181exeFLOPs
181exaFLOPS / 40GFLOPS = 2^n, n=32
32*18mo = 48 years based on personal computer technology

or 28 years based on supercomputer technology

of course a real neural network will contain highly parallel processing and using a specific chip design we will probably be able to simulate a brain much sooner, perhaps in the order of 10-20 years.