Slashdot Mirror

...he says that GISS refused to give him the algorithm's source code.

Given that this is a government organization, I think the refusal to hand over the source code is wrong (obviously there is no national security concern). Could a FOIA request force this?

For, scientists working outside the government, I think providing the algorithm and data are sufficient, although even in those cases I would prefer that they open source (not necessarily GPL, though) their code. (We have.)

The first place to go to get ideas about applications that can be sufficiently parallelized is to consider all of the @Home programs. In addition to those, there are the ones already mentioned by those who have responded to you. Then there's my research area - neural network models of mammalian brains. One of the things that people plan on doing on Blue Gene is simulating an entire human brain.

We have an open source project that models brain regions, that is extremely unlikely to ever be widely used by a general audience. However, if it were used by 25% of neuroscientists who run brain simulations, I'm sure we'd consider it successful.

I have not been personally involved in any detailed models of neurons, but there are models out there of single neurons with thousands of compartments. These definitely don't fit the "black box" description, IMO, and I think they do quite a good job of modeling the behavior of the neurons they're designed to model. I don't know if one has been worked out for a cortical pyramidal cell, but I do know they work quite well for hippocampal pyramidal cells, and I'd be very surprised if they hadn't been worked out for their cortical brethren.

One strategy we've explored is a merging of those detailed models with our simplified models. We've created a rudimentary interface that allows our neural network software package to write out data such that a Neuron script can read it in, thus allowing comparisons between the two levels of detail. We can also reverse the process (have our software package read in input from a Neuron simulation). What I could envision happening is that one uses GAs (or some other approach) to iron out some of the details of the detailed model (as the GP experimentalist admits there is unfortunately still a lot of unknowns about the in vivo behavior), and then use another GA (or Evolutionary Algorithm) to find a suitable simplified model of the detailed model - one that is simple enough to be run in a simulation of 100,000 neurons.

I've taken several courses in thermodynamics and several courses in statistical mechanics. After all, I also have a degree in astrophysics, and I can honestly tell you that this isn't rocket science, either. ;) Simply put, I understand statistical variance quite well. However, deterministic processes are supposed to be deterministic. Sure, it's possible that a bit in a computer will randomly flip (hence the reference to neutrinos, which are usually the humorously blamed party), but if bits flipped that often, then I'd NEVER be able to get the same results in my program that requires more than a billion synaptic events. (Have I not mentioned this program before?)

Let me put it to you this way: if I drop an egg a million times, will you argue that statistics mandates that there's no way the egg will fall every time?

I have a "quick test" that simulates billions of synaptic events. This "quick test" must be passed every time before I check in any changes to the program running the quick test. I am not satisfied if a single synaptic event is missed, unless I understand the reason that one (out of over a billion) synaptic event is different.

Barring fraud, why do assert that "Voting machines (or processes) will never reliably count one hundred million votes"? It's not exactly brain surgery (pun intended).

If I add more cores, it's so I can simulate bigger neural networks in close to real time. Depending on my level-of-detail, 10,000 neurons take anywhere from 250-750 MB. In order to run 100,000 neurons I need dozens of CPUs (currently done via Beowulf clusters), and each CPU needs a lot of RAM. (Up to about 100k neurons, synapses scale roughly as n^2, not n - at least in our model of the hippocampus.)

In yet another article in Ars Technica we read that Intel is look to an 80 core chip. I like the Core 2 Duo a lot but I hope the Intel megahertz fixation isn't just going to become a "core" fixation .

Of course, as others have said with Citizendium et al., it still makes an excellent addition to the wiki-verse. Well, at least for those of us with an interest in computational neuroscience, dynamical systems, and/or computational intelligence. Presumably, additional scholastic areas will be added as demand calls for it. (If you are interested in these areas, you might also enjoy NeuroJet.net. It, too, has a wiki, but it's only helpful if you're planning on using the neural network simulator that is NeuroJet.)

I also want to give a shout out to Eugene Izhikevich (founder and editor-in-chief of Scholarpedia) for his contributions to all of those fields.

You keep using the phrase "left winged 'scientist'". I do not believe the phrase means what you think it means. Having an MS in Physics/Astronomy, another Masters in CS, and working on a Ph.D. in Computer Science in a sub-field that involves a lot of neuroscience (i.e., practically being a professional student), I've been exposed to many, many people that I'd qualify as scientists. Every single one of them, to the best of my knowledge fits your definition of "left winged 'scientist'" even though many of them (but a minority) voted for Bush in the last election (don't ask me why). Of the hundreds of college students I've known, only 2 of them would not qualify as being "left winged", and I've matriculated at Georgia Tech (BS/Phys), Georgia State (MS Phys/Astr), and the University of Virginia (MCS, Ph.D. in CS in progress). (These are three colleges that are relatively conservative.)