From: Damien Broderick (d.broderick@english.unimelb.edu.au)
Date: Tue Aug 27 2002 - 01:52:15 MDT
Is he a dud? Why no, he's now an Associate Prof at Utah.
http://www.cs.usu.edu/~degaris/
And the following grumpy tale sounds a bit like the start of Heinlein's
DOOR INTO SUMMER:
===============
http://www.cs.usu.edu/~degaris/news/CBMdown.html
What Happened to the "CAM-Brain Machines" (CBMs)?
After all the world wide publicity I was getting in 2000/2001 concerning
the creation of the (4)
CAM-Brain Machines (CBMs), (which were capable of evolving a 1000 neuron
neural network circuit
module in a few seconds, and then updating the neural signaling of 64000 of
these modules
interconnected in a gigabyte of RAM, in real time (constituting an
artificial brain of nearly 100 million
neurons)), a lot of people have been asking me "What happened to the CBMs?"
In a nutshell, the
CAM-Brain Project died prematurely, due to the bankruptcy of Starlab
(Brussels, Belgium, Europe),
where I was working from February 2000 to June 2001.
Starlab bought a CBM early 2000, which was delivered in the summer of 2000
- price tag, $0.5M. From
the beginning, we had problems with it due to inadequate air-conditioning
in the older Starlab building
(before the move in November 2001 to the much bigger building), and
frequently when the CBM was
turned on, the power in the building would switch off. The CEO of Starlab
presumably then got the
impression the CBM was a dud. This was not true, because as soon as the CBM
was transferred to the
larger building, with proper air-conditioning and proper electrical wiring,
it worked fine, but by that stage
the CEO seemed to be too stressed/preoccupied to be reading my emails that
the machine was working.
Every time I mentioned to him that he had a contract with my HW colleague,
he would explode like a
petulant child. My assurances to him that the CBM was working and meeting
specifications, had no
effect.
The CEO did not pay my hardware colleague all of the money he had
contracted to buy the CBM. My
hardware colleague finally sent a very threatening email to the CEO which
put his back up totally. Later,
the hardware colleague switched off the CBM at Starlab. (He had internet
access to the machine to update
regularly its firmware). This killed the CBM research project. My hardware
colleague cared more for his
money than the continuation of the brain building project. This alienated
me from him, so that the two of
us no longer even email each other any more.
I had $100,000 of my own money invested in Starlab, i.e. the CEO, so I was
hardly going to "bang my
fist on the table of the CEO" to get him to pay my HW colleague (the
designer of the CBM, based on my
ideas). The CEO hated my HW colleague, whom I also found far too brutal.
Towards the middle of 2001, it looked as though Starlab would get a bridge
loan, some of which would
be used to pay off the CBM, so that my HW colleague would switch on the
machine again. I went off to
my summer conferences believing that the CBM payment problem would be
solved on my return. What
happened in fact, was that the lab went bankrupt, so my hopes of seeing the
CBM up and running were
dashed. In fact I only saw it running for about a month. (It really did
evolve neural circuit modules in a
few seconds. It was breath taking).
My main challenge now (summer 2002) is to find the money to continue this
effort, but this time, the
hardware design expertise is now in-house. I and my research assistants and
other collaborators have
become hardware programmers ourselves (using programmable chips, mainly
from Xilinx). I never want
to be in such a position of dependency on another person, such as my ex HW
colleague, who could kill a
decade's work by switching off my access to the machine.
So, if future years show that brain building is a success, then the failure
of this first generation brain
building machine attempt might be of interest to the historians of technology.
What happened to the 4 CBMs around the world? With my ex HW colleague no
longer working on the
project, there has been no progress on the machines, so they are probably
gathering dust in labs in Japan
(1 machine) and Europe (2 machines). The ex HW colleague has one of his own
(US), so perhaps he is
playing with it.
I have rather bitter feelings towards both my ex CEO and my ex HW
colleague. I felt they were both
responsible for the tragedy of the demise of the CBM. I was not in a good
state at the time. My 2nd wife
died January of 2000, so I wasnt fully functional either. Life can be tough
sometimes. But, thank god for
Moore's law. Even if the CBM were a great success, I would still have to
build a new machine every 5
years, so by the time the 3rd or 4th generation machine is built in the
future, the fate of the 1st generation
machine will be irrelevant. The basic idea of building artificial brains
using evolvable/programmable
hardware remains very valid. It will happen. It could have happened in
2001/2002. Now it will have to be
postponed for a few years while the second generation machine gets funded,
designed, built and tested.
=======================
That damned receding Spike... But wait, he's now working on quantum
parallelism:
=======================
I and my research assistants are now working on quantum algorithms applied
to neural net creation. In fact, our group no longer speaks of neural net
evolution, since evolutionary algorithms are only needed in a classical
computing context, where a classical computer can only process one Nbit
bitstring at a time. Since there are too many (2**N) possible chromosomes
in the search space, only a sampled search can be undertaken, which is what
an evolutionary algorithm is. BUT, with quantum algorithmic techniques, all
2**N chromoses can be processed at once, thus killing the need for
evolutionary algorithms. Our group has discovered quantum algorithms that
measure the fitness of all 2**N chromosomes (hence quantum neural networks)
at once. Most quantum evolutionary computation approaches use a classical
genetic algorithm and apply it to generate a sequence of quantum operators
operating on a string of quantum bits (qubits), which is a form of
classical/quantum hybrid. We use a fully quantum approach, and thus dont
need an EA, since the whole 2**N point search space is explored
simultaneously. Thus we speak of quantum optimization rather than quantum
evolution. In this case, QC has made EC redundant. My prediction is that QC
will largely kill off EC in the future. My opinions on this point are still
evolving ;-)
================================
Damien Broderick
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:16:26 MST