Slashdot Mirror

You found the paper, but you didn't look at any of the followup research.

Like this paper which details an experiment using an external clock and a wide variation in temperatures to evolve the same sort of circuit that Adrian evolved in his thesis paper.

And a complete list of his publications can be found here.

If you've bothered to read any of his work, you'd quickly realize that Adrian is interested in how evolution can use certain properties of the physical substrate in these chips to it's advantage. It's not looking to see if evolutionary type strategies can evolve something a human could build, but looking at how they can build things no human could imagine building.

DISCLAIMER: I am currently a Master's student at the University of Sussex, and had Adrian as a lecturer this past semester. However, I am in no way involved in his research, my interests lie in the software side of genetic algorithms.

I agree, Thompson seems to have been doing some really wierd stuff recently. I mean, single electon gate design with genetic algorithms?

I found the paper on this project, and I found a few things disturbing. First of all, there was no clock: the circuit was completely asynchronous. In other words, the only timing reference they had was the timing of the FPGA itself. Trying to do something like this in silicon is difficult, and doing it in an FPGA is just plain insane. Delays in a circuit vary with just about everything: power supply voltage (and noise), temperature, different chips, the current state of the circuit, and so on. While you might be able to deal with these problems in a custom chip, an FPGA was never designed to be stable in these respects. Also mentioned is that there are several cells in the circuit that appear to have no real use, but when removed, the circuit ceases to operate. As they mention, this could be because of electromagnetic coupling or coupling through the power supplies. Again, I would never want to see something like that in one of my chips.

Another thing that bothers me, how the heck does he know which cells are being used? Last time I checked, the bitstream (programming) files for these chips is extremely proprietary, and nobody (except XILINX) has the formats for these files. I really want to know how they know how this thing is wired.

Now I should mention, this is pretty cool from an academic standpoint, and it would be interesting if they could produce something that is both stable and useful using these techniques. It's also pretty cool that they could get this to work at all.

While i have to say that Intel's OpenCV library rocks (for a number of reasons), stereoscopic vision is nothing new. The cnn article is more or less crap ("Until today, computer vision applications has been restricted to two dimensions
"? nice try...) It's mishmash of reporter hype and stock text which describes computer vision in general ("Over the next 5 to 10 years, Intel Corp. expects computer vision to play a significant role in simplifying the interaction between users and computers"). The Sussex Computer Vision Teach Files page has a reasonable description of stereoscopic vision from 1994. Lip reading is not really a 3D problem, so stereoscopic capabilites aren't going to help much. Many of the other uses- 3D environment modeling, object modeling and recognition, etc, are being worked on (again, the algorithms aren't new, this is just a new open source implentation) but they're not easy.

I don't mean to sound pessimistic, though. OpenCV is really cool, both as a corporate contribution to open source, and as a programming library even if you never look at the code. And the Matlab interface means fewer MSVC++ sessions which end with me feeling homicidal ;-) The inclusion of stereo vision will be cool for people trying to write vision applications, but it's not advancing the state of the art.

While i have to say that Intel's OpenCV library rocks (for a number of reasons), stereoscopic vision is nothing new. The cnn article is more or less crap ("Until today, computer vision applications has been restricted to two dimensions
"? nice try...) It's mishmash of reporter hype and stock text which describes computer vision in general ("Over the next 5 to 10 years, Intel Corp. expects computer vision to play a significant role in simplifying the interaction between users and computers"). The Sussex Computer Vision Teach Files page has a reasonable description of stereoscopic vision from 1994. Lip reading is not really a 3D problem, so stereoscopic capabilites aren't going to help much. Many of the other uses- 3D environment modeling, object modeling and recognition, etc, are being worked on (again, the algorithms aren't new, this is just a new open source implentation) but they're not easy.

I don't mean to sound pessimistic, though. OpenCV is really cool, both as a corporate contribution to open source, and as a programming library even if you never look at the code. And the Matlab interface means fewer MSVC++ sessions which end with me feeling homicidal ;-) The inclusion of stereo vision will be cool for people trying to write vision applications, but it's not advancing the state of the art.

> They can't hide devious little "big brother" bits of code in their OS because we get the code with the OS.

You've got the source code; have you personally audited it from top to bottom, and verified that there are no back doors? Even if you have, have you also made your own compiler to rebuild everything?
If you've ever heard the phrase "security by obscurity", don't pretend that openness is a magic bullet; just because you and a herd of others bury your head in the sand doesn't mean that back doors aren't already in there.
For what it's worth, MS releases code to large clients; if there were glaring holes in there, well. I'd say they'd be released to the public, but I wouldn't doubt they'd pull NDAs to cover their asses a la Sun...

Better write them down in covariant form - even shorter and more elegant, and more geeky (will not be recognized by EE students however).

Here's an interesting link discussing this type of reactor.

...or rather here.

... here.

here

Here we go. Gribbin talks about it in Schroedinger's Kittens, but also discusses it on his web page.

Hope this helps....

1. This book seems a little over-excited or over-hyped in the review. We haven't created anything too impressive yet.
2. Alife is just about the coolest thing ever. What we are hoping to make is pretty exciting.
3. Yes, definitions of life can cause problems, but what is new about ALife is that it tries to approach everything by learning from life, rather than just introspection about how we might think or behave, which is what traditional AI is based on. It also generates artefacts that are interesting to some people because they show characteristics that had previously been thought to have been exclusively shown by "real" lifeforms.
4. I think the Turing test is totally inappropriate, since it seems to be neither necessary nor sufficient for life. If you talked for hours to a machine through a wall and believed it was a person, and then I showed you the machine, would you think it was alive?
   (...clearly it's not a necessary condition.)
5. ALife is being applied to loads of stuff, from very abstract, through scientific, engineering/robot design to entertainment. Check out http://www.cogs.susx.ac.uk/users/ezequiel/alife-pa ge/alife.html for a fairly comprehensive list of the more academic stuff.
6. I don't agree that consciousness is the vital ingredient. In fact, all I look for in a "good" experiment is interestingness.

Have a look at www.artificialworlds.net for some fun with ALife (and a touch of AI).

this is because proteins that genes produce can do more than just go away and make stuff for bodily processes, they can affect (turn off, turn on, conditonally flip, etc) the action of other genes.

so many people seem to have this idea that genes are just like a recipe, a linear list of stuff that goes to make up the organism. they're nothing like that.

your genes are a program like the worst spaghetti code ever written, multiplied a billion-fold. every emergent property of the system will be used - there's nothing to tell evolution to Keep It Simple, Stupid.

where does that leave the Human Genome Project? well, they know *something*. but my bet is that there are fundamental limits on the amount that we can ever actually know about how our genes work.

check out the papers on this page for a fascinating example of how evolution can create things we don't understand, even in an extremely limited domain.

also, Stuart Kauffman (of the Santa Fe Institute) has written a book "at home in the universe", the first part of which gives an excellent exposition of the above boolean net stuff. highly recommended.

anyway, isn't it obvious that the complexity of a system bears no relationship the complexity of the rules of that system. think of Life for example: only three rules, but given a large enough arena to play in, those rules are sufficient to simulate any computer... start it off from a random state, and it's fundamentally impossible to predict what it's going to do!

http://www.cogs.susx.ac.uk/users/adrianth/ade.html

Nearly always, the circuits that evolve are smaller than those that are designed by an engineer. Here, the gates of an FPGA appear analogous to genes. It seems that these experiments with FPGA's might have predicted fewer human genes. I wonder what else these experiments might predict is going on in humans.

These FPGA's have odd failure modes. Is there a connection with certain human diseases?

Adrian Thompson has been researching hardware evolution; using genetic algorithm as a feedback loop for programming FPGAs. In some cases the optimal solutions achieved obviously relies on complex electromagnetical resonance from seemingly unconnected parts of the circuit, behaviour not anticipated by todays testing suites.

Also, if by regularity of digital circuits you mean its deterministic logic, it's ability to not be affected by random fluctuations of its environment, I'd like to point out that that might not be a positive property in this context.

> it's perfectly happy bonding with itself

Me too! :-)

But seriously, I did some undergrad work with high temp *super* conducting C60 intercalates - using things like potasium and sodium to fill the gaps in the C60 crystal structure... Groovy stuff with a lot of bright people working on it.
Here is a link to the Sussex fullerene research centre....

If they can pull this off - it would be somat!

The groups leader Prof Kroto jointly won the Nobel gong in '96

Of course the holy grail is *room* temp super conductors.
High temperature (at the moment) wrt super conductors is about 40K!

i think it's fascinating that here's a circuit that they built themselves, and they still cannot figure out how it works... reverse engineering has its limits!

PS. isn't code stored in NVRAM and used to program reconfigurable logic chips subject to the same copyright laws as any computer program? so presumably ripping this off and using it in your own product has about the same legality as selling a dodgy copy of MS Word... despite what others have said about trade secrets, there seems to be a fine line here between a "trade secret" and a "copyrightable piece of code". hmm.

there's a major problem there: the assumption that there's a reasonably straightforward mapping from gene to meme. but i think that current assumptions are probably highly simplistic, and in fact, perhaps in general the mapping is not discoverable.

not so long ago, i read most of a book by Stuart Kauffman called "At home in the universe: the search for the laws of self-organisation & complexity". It became a little less solid towards the end, but the early chapters contained what was, to me, a brilliantly illuminating discussion of the way that networks of genes function.

kauffman argues that the expression of a genome is not just just the simple reading of segments of the DNA which encode proteins which go away and build things, but that a set of genes really forms a boolean network, where the action of some gene can affect the expression of another gene, and vice versa.

what that means in computery terms is that the way your genes work is less like a shopping list (gene A implies obesity, gene B implies intelligence, etc), and more like a cellular automata. if you remember some of your computer theory, you might remember that many simple CA's (e.g. Conway's Life) are Turing complete.

so what we have is essentially an evolved computer program. and if you think that some people write bizarre code, wait till you've seen some that's generated by genetic algorithm. then multiply that by billions of year's worth of evolution, raise to the power of the Halting Problem, and that's the order of the difficulty of decoding the genome!

by way of illustration of the sort of complexity that can arise when even simple systems are evolved in the real world, check out Adrian Thompson's web page. In particular, this paper has a fascinating analysis of the properties of some genetically evolved FPGA hardware. now this stuff is really simple - we're talking digital components, 100 gates, evolved to perform a simple discrimination process.

the circuit worked, but they didn't really have the faintest clue of how! because it evolved, it pushed the physics of the FPGA as far as they would go. to quote from the paper:

There are numerous tactics that can be used to piece-together answers to analysis questions even for seemingly impenetrably circuits. We applied many of those techniques to the most advanced unconventional circuit yet produced. We still do not understand fully how it works: the core of the timing mechanism is a subtle property of the VLSI medium. We have ruled out most possibilities: circuit activity (including glitch-transients and beat frequencies), metastability, and thermal time-constants from self-heating. Whatever this small effect, we understand that it is amplified by alterations in bistable and transient dynamics of oscillatory loops, and in detail how this is used to derive an orderly near-optimal output. Certain peripheral cells fine-tune particularly sensitive time delays.

that's why i argue for caution in the use of genetic engineering technology. actually, i'm not sure i do. nature has thrown so many genes together for so long that i doubt we can come up with much that does anything really useful that isn't just a simple isolated gene-to-attribute mapping.

the claims that are made for genetic engineering are way overblown - genes might be the roadmap for life, but i bet they'll be an almost completely unreadable one.

there's a major problem there: the assumption that there's a reasonably straightforward mapping from gene to meme. but i think that current assumptions are probably highly simplistic, and in fact, perhaps in general the mapping is not discoverable.

not so long ago, i read most of a book by Stuart Kauffman called "At home in the universe: the search for the laws of self-organisation & complexity". It became a little less solid towards the end, but the early chapters contained what was, to me, a brilliantly illuminating discussion of the way that networks of genes function.

kauffman argues that the expression of a genome is not just just the simple reading of segments of the DNA which encode proteins which go away and build things, but that a set of genes really forms a boolean network, where the action of some gene can affect the expression of another gene, and vice versa.

what that means in computery terms is that the way your genes work is less like a shopping list (gene A implies obesity, gene B implies intelligence, etc), and more like a cellular automata. if you remember some of your computer theory, you might remember that many simple CA's (e.g. Conway's Life) are Turing complete.

so what we have is essentially an evolved computer program. and if you think that some people write bizarre code, wait till you've seen some that's generated by genetic algorithm. then multiply that by billions of year's worth of evolution, raise to the power of the Halting Problem, and that's the order of the difficulty of decoding the genome!

by way of illustration of the sort of complexity that can arise when even simple systems are evolved in the real world, check out Adrian Thompson's web page. In particular, this paper has a fascinating analysis of the properties of some genetically evolved FPGA hardware. now this stuff is really simple - we're talking digital components, 100 gates, evolved to perform a simple discrimination process.

the circuit worked, but they didn't really have the faintest clue of how! because it evolved, it pushed the physics of the FPGA as far as they would go. to quote from the paper:

There are numerous tactics that can be used to piece-together answers to analysis questions even for seemingly impenetrably circuits. We applied many of those techniques to the most advanced unconventional circuit yet produced. We still do not understand fully how it works: the core of the timing mechanism is a subtle property of the VLSI medium. We have ruled out most possibilities: circuit activity (including glitch-transients and beat frequencies), metastability, and thermal time-constants from self-heating. Whatever this small effect, we understand that it is amplified by alterations in bistable and transient dynamics of oscillatory loops, and in detail how this is used to derive an orderly near-optimal output. Certain peripheral cells fine-tune particularly sensitive time delays.

that's why i argue for caution in the use of genetic engineering technology. actually, i'm not sure i do. nature has thrown so many genes together for so long that i doubt we can come up with much that does anything really useful that isn't just a simple isolated gene-to-attribute mapping.

the claims that are made for genetic engineering are way overblown - genes might be the roadmap for life, but i bet they'll be an almost completely unreadable one.

The researcher I believe you are looking for's name is Adrian Thompson. His web page is here. There is also an article on Discover's web site, if you go to their archives section and search for "FPGA" in the _body_ of the article. The article is called "Evolving a concious machine" and is by Gary Taubes. (Surprisingly it is the only article that contains the word FPGA in its body!)

I haven't looked at his work in a while, but I'm sure he has done some cool things with his evolving hardware since 1998. I always thought that the most interesting part was that he didn't limit the evolution to digital-only solutions-- resulting in incredibly efficient circuit designs that make *use* of crosstalk and interference!

From what I understand though JK's description, Mr Harel is probably talking about the NP-hard problems, ie problems which take exponential time to solve (exponential being related to their "size", eg solving the travelling salesman problem for N cities takes k*exp(N) steps).

Although those problems are effectively unsolvable through the classical, algorithmic way, quite a lot of them can be solved using the most recent AI techniques - the drawback being that the solution is not 100% guaranteed optimal. Genetic Algorithms [?], for example, are the most powerful optimization tool that ever came out of AI. It can deal with the travalling salesman's problem (see one version here), just as well as other technique such as "Ant colonies"

Furthermore, complexity theory (which deals with "computability") only holds for Turing machines. DNA / quantum computers do not fell in the "NP-cursed" category of computers.

Mr Harel's thoughts, while being perfectly snesible as far as his own field is concerned (Turing-like algorithmics), should not be taken as holy scripture. Digital calculators are only a couple of decades old. It took thousands of years to fully exploit the power of the steam engine. We can try to imagine what "computers" will be like in 30 years from now, but expecting such a forecast to be accurate would be foolish.

Thomas Miconi

Looks like Thompson's still working on exploiting the non-digital properties of digital devices, if I understand the blurb.

Looks like Thompson's still working on exploiting the non-digital properties of digital devices, if I understand the blurb.