Slashdot Mirror
Electronic Circuit Mimics Brain Activity
A lot of people wrote in with the news blurb from Yahoo! regarding the announcement of a ciruit that supposedly acts in a manner resembling human brain activity. Details in the blurb are pretty sketchy though - post links below if ya got 'em. One of the interesting points that they say though is that the brain does both digital and analog - but that's pretty much all they say about it.
/joeyo
2^5
I'd say that the main flaw in Penrose's argument stems from the fact that he seems to be seeking religion rather than searching for scientific (ie. testable) explanations.
I like well-written books that propose alternative theories, but they've got to have some sort of solid framework and internal consistency to be worth reading. The Emperor's New Mind was great as long as Penrose stuck to reviewing previous science, but appalling thereafter. I don't recall ever having read a popular science book containing so much handwaving, copouts, and defeatism. He's desperate to prove that scientific investigation is dead in the water when it comes to the mind, it seems to me. Put that together with some of the mystical mumbo jumbo that appeared liberally and it all starts to add up to a personal search for his God and The Reason He Must Exist.
Bleh, a very disappointing read.
"The question of whether machines can think is no more interesting than [] whether submarines can swim" - Dijkstra
The relation between firing and other neurons noticing and caring is not digital. However the neuron has binary way of attempting to communicate - namely firing.
Cheers,
Ben
My usual seat in the cluetrain is at A HREF="http://pub4.ezboard.com/biwethey.ht
FWIW my understanding of neurons is based on conversations with my wife who happens to have a PhD in biology and is pursuing her MD. (Note, the combined PhD/MD is considered a weaker combination than separate "real" degrees.)
Now I grant that neural networks may be different. And there may be differences between neurons.
But I definitely know that your basic neuron sends a stronger signal by firing more often, not by firing more strongly. Ditto for nerves and sensation. Stronger sensations are caused by more rapid firing, not more intense firing.
Regards,
Ben
My usual seat in the cluetrain is at A HREF="http://pub4.ezboard.com/biwethey.ht
A neuron's life comes down to deciding when to fire. Fire/not fire is a binary decision. There are not different types of firing. You do or you don't.
OTOH the neuron's decision to fire is influenced by all sorts of things from the chemical balance, what other neurons have fired recently, whether it tends to fire with them, etc.
So a neuron makes a digital decision on analog criteria...
Cheers,
Ben
My usual seat in the cluetrain is at A HREF="http://pub4.ezboard.com/biwethey.ht
Sadly, work on neural networks still sometimes relies a lot on buzz. Nature, as a journal, seems particularly susceptible to this kind of science: what they publish has to be short and pithy.
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!
Penrose's "The Emperor's New Mind" and "Shadows of the Mind" also makes the case (quite effectively, imo, but you may disagree) that the mind is not a digital computer, but a quantum computer and that to get a computer to think like we do we'll have to make it model quantum effects.
Admit nothing, deny everything and make counter-accusations.
The main flaw in Penrose's argument is that he gives no mechanism for the human brain to exploit quantum computation.
Of course, there are many other flaws, such as his assumption that humans aren't susceptible to "godelization". Sure we aren't susceptible to the SAME godel strings that number theory and turing machine are--that doesn't mean we are perfect.
Check out "Fabric of Reality" for a little more on this. There was another book I read recently that rebutted Penrose more effectively (and more thoroughly), unforunately I don't remember what it was.
--
Linux MAPI Server!
http://www.openone.com/software/MailOne/
(Exchange Migration HOWTO coming soon)
>I don't agree 100% that the brain makes digital decisions.
Not at the level of conscious thought, that is for sure.
However, at the level of the individual neurons, the response to a stimulus (whether to fire an action potential and at what frequency) are pretty much determined by the configuration of the cells and the electrochemistry of the cell membrane.
I think that the article just didn't make this clear.
LL
"If you are falling, dive." -Joseph Campbell
User:But...
Computer:Oh, and another thing, I was lying. I've seen much bigger hard drives.
Dreamweaver
"If a man hasn't discovered something he will die for, he isn't fit to live" -- MLK, Jr.
Somebody else made an excellent post describing this stuff in more detail.
To sum it up, the neuron acts pseudo-digitally. It must first determine if the stimulus is enough to fire. But once it has determined it is, it then fires an analog signal, the power of which is encoded by the firing rate.
Contrast this to a purely digital neuron, who, after recieving a stimulus, will just fire normally, and not really give any analog data as to how powerful the original stimulus was.
So real neurons are sort of passing some more info along, and I assume this allows for all sorts of subtle and nuanced feedback loops, etc., that may not be possible in a completely digital neural net.
It's 10 PM. Do you know if you're un-American?
Well, my impression is that if some level of stimulation hits an analog neuron, that neuron can fire others with some fraction of that stimulation. While a "digital" neuron would have to determine whether the signal was "enough" stimulation and if so, stimulate the others, otherwise don't stimulate them at all.
It's 10 PM. Do you know if you're un-American?
This was my first thought too. While the brain may make either-or decisions, that has no bearing on the actual nature of the process. Analog circuits can easily make "digital" decisions.
I think the problem lies in the author of the article. According to another post, the analog-digital thing happens on a neuron level. So the Yahoo article's explanation is just a bunch of hooey thrown in for those that won't question it.
How many times must we read about this kind of thing? We already know where it leads to: 1. Electronic circuit built to mimic human brain. 2. Circuit is put into super powerful Computer. 3. Computer reaches self-awareness and self-actualization. 4. Humans forced into servitude to the Computer. When will they ever learn?!?!- ----
--------------------------------------
Anyway, when he was done he had an FPGA that could tell the difference between "Stop" and "Go". The interesting part was that the program that it used wouldn't work on other FPGAs. Apparently, it was using analog effects that were specific to the individual chip. Furthermore, it was really efficient. Only a small percentage of the chip was being used. (Does this sound like your brain at all to anyone else?)
I was wondering if anybody had heard anything more about this research. I think it is facsinating.
First, we have a way to control smartness in animals, then we have a way to make electronics act like a brain. Combine those, and set the whole thing in flowers for algernon:
Janyuary 2023
My yuser think I stoppid, but I now I not be stoopid. How can I be stoopid wen I rite al these algo- alga- algarithims I think theyre caled. Yesterdy I rite a BSOD and my yuser no lik it. He say a nice peeple will help me and make me gen-yus. Just like Liynux. He is a mice with a jene to make him gen-yus. They saying they will do this too I. I hop I became gen-yus just lik he!
nuclear cia fbi spy password code encrypt president bomb
Friends don't let friends misuse the subjunctive.
The short answer is "We don't entirely know yet." The longer answer is way too complicated for me to get into in a Slashdot comment; I strongly suggest that you read an intro-level neuroscience or cognitive psychology textbook, or at least a single chapter therein. The basic theory is that as you learn things, your neurons physically change shape and alter their connections, and information is encoded in these connections.
Also, be very careful of things like that bicycle accident memory. Vivid images like that are called "flashbulb memories", and studies have shown that they tend to be inaccurate as hell. In fact, the more details you think you remember about a single instantaneous event, the more likely you seem to be to be wrong.
There are many different types of memory, and they appear to be stored in different brain locations with different encodings. These mechanisms are under heavy study, but are not well-understood. Trying to talk about the "storage capacity" in terms of megabytes is essentially futile at this point in time.
Look, the important thing is not that we mimic the human mind or human thinking. Why do we want machines that think and act like humans? What good is that? So we can understand ourselves? Well, that is silly since the mechanisms that drive our intelligence are simply not going to be the same as the machines we make with human intelligence. That is, a computer with human intelligence tells us nothing about what really makes human intelligence actually work. The best a machine can do is ghost our cognitive economy, it cannot actually have it.
But that might be beside the point. More important is that we build and understand machines that have a higher level of intelligence than us. That intelligence might be nothing like a human's intelligence, but that's fine. As we all know, computers have a different kind of intelligence than us. And that is interesting. That should spark our creativity and that should get our juices going.
Here's an analogy. Suppose I build a telephone out of rubber bands and paper clips. It acts just like your favorite phone. But, is that interesting really? I mean, is the fact that we have a "really cool copy of a phone" all that interesting in terms of what-it-is-to-be-a-phone? Of course not. Instead, it is interesting that the damn thing is so complex and useful, even though it was made from rubber bands and paper clips.
Forget mocking the human experience. We get that each day, don't we? We get it (we're human). Let's look at other kinds of intelligences, based on machine mechanisms.
John S. Rhodes
WebWord.com -- Industrial Strength Usability
How to Download YouTube Videos
This invention is really a small step in the direction of having computers mimic the brain's capabilities at some cognitive abilities. For example, recently IBM showed that Big Blue, a computer, could beat the world's best at chess. The brain still has many areas at which it cannot be beat. Such as
Pattern recognition with translational invariance, rotation invariance, and size invariance
Speech recognition in noise.
Having computers that could perform things like pattern recognition or speech recognition as well as humans would allow enormous advances in the roles of humans and computers in our lives. People like Sebastian Seung think inventions like this will take them down that route - and ultimately result in huge scientific advances in artificial intelligence.
Personally, I think studying how the BRAIN does pattern recognition will allow far faster advances in this area than inventing chips that have SOME of the capabilities of neurons.
Neuromorphic electronic elements have been around a long time. There was an article from the 60's (in the Cold Spring Harbor symposia series, if I remember correctly) describing a vacuum-tube implementation of a neuron. I thought that was hilarious, given the hype that was going into the more recent silicon versions There is a basic design principle: don't do in hardware what you can do in software. Neuronal simulator packages like GENESIS and NEURON can simulate extremely realistic neuronal models in real time or faster. These neuronal models use 'compartments' which are little cylindrical segments of the cell, and they build up the geometry and electrical properties by putting them together. In other words, it is a spatial discretization of the cell. The finer the divisions, the more accurate the model. Last time I benchmarked, a PIII or Athlon class PC could handle something like 100 compartments in real time. That could either be 10 rather coarsely modeled neurons, or 1 very accurately modeled neuron. It is a great deal easier and cheaper to throw several PCs together in parallel to make a big network model, than it is to design an analog VLSI chip from scratch. The models are much more flexible, and can incorporate the latest data. You don't even want to think about what the debugging of VLSI would be like... I think Carver Mead's approach was the most practical: use the principles from real neural circuits, embody the equivalent computations in analog VLSI without being too picky about how neuron-like they were, and really use VLSI on a large scale. That is what he did with the artificial retina.
Anyone heard of the chaos theory here? Of course you have. Then you'll know that quantuum effects will propagate through all reality, not just even itself out over time. _Especially_ in the real world. (If you put things in a simulator, things _may_ converge on a grander scale, or not. It depends on the rules for feedback, what start-up conditions you begin with and what kind of number-system you use.)
So quantuum effects will always have _some_ effect. However, if it's big enough for dramatically changing how we think is another question. Alas, the whole sherade might not be so tied with our brain as we'd like to think either. Higher processing may manifest itself in quantuum effects in everything around us, including our whole body.
The problem is proving all this. Thank god everything can't be proved.
- Steeltoe
http://www.debunkingskeptics.com/
Regular neural networks still work on digital information only. These things, apparently, do not. That's why it's a big deal.
I do have a problem with this statement in the Wired article though:
Claiming that these guys have pioneered mixed-signal design is just a little bit of a stretch. Do your research, Wired. =)
--
Given a hand, a pencil, an eraser, and an infinitely long piece of paper, a brain can easily emulate a Turing machine. Does that make it a Universal Turing machine? (I think emulation is the only criterion). Even without the paper and pencil, it can emulate such a machine, apart from the poor storage capability.
Well, welcome to the real world! It's not always fair and sometimes people play unfairly to gain an advantage. What, do you think the media just snoops around MIT constantly looking for stories like this? MIT most likely has a good PR department (they'd be stupid not to).
Did the U of Manitoba do a press release on what they were doing? I didn't see it. Plus, this is apparently a joint corporate/university operation, so that's probably another reason we see it in the news.
Oh, and I don't mean to be cynical, but just because it happened doesn't mean you'll see it in the news -- another reality check for you.
As computer clocks go higher and higher, designers are going to have to become more and more aware that there's no such thing as a digital circuit. All electric and electronic circuits are analog.
When's the last time a computer designer had to worry about impedance matching between his circuit board and the components on it? As circuits become smaller and smaller electrically, transmission line effects become more and more important. Suddenly the digital designer finds that absolutely none of his signals are making it past the package leads due to lead inductance, or the dielectric constant of that cheap plastic package is high enough to cause the characteristic impedance of the line to be ten times lower than the PC board trace!
At least as an RF circuit engineer, my career is secure :)
Anyone who thinks that the human brain is a Turing Machine cannot consistently believe this sentence to be true.
-- the most controversial site on the Web
The great evolutionary leap to be done is in the field of 'learning algorithms' that can be applied to neural nets with lots of feedback (the output of one neuron is fed to a neuron which is in a previous layer - farther from the output).
Unfortunately, the article says nothing about this. If they've just created a hardware neural net, Duh! It doesn't even deserve a footnote.
If at first you don't succeed, skydiving is not for you
Biological (neural) systems have properties sometimes desirable electronically, such as robustness and insensitivity to noisy data. Indeed, Caltech's Carver Mead (if he's still there) went a long way to popularize biologically-inspired engineering, or "neuromorphic engineering." His book Analog VLSI and Neural Systems is the usual text, mixing VLSI design and mimicry of, say, the retina.
The original Nature article should be readable to those clued in on MOS circuitry and a bit of neuroscience. I think it's wonderful that Nature is willing to post their material for free online, esp. in PDF...
For those of you itching to learn more about the brain & neuromorphic engineering, I set up a page of links to related books.
All best,
Gregg Favalora, CTO, Actuality Systems, Inc.
Developing autostereoscopic volumetric 3-D displays.
Wired News offers a little more detail. The expressed difference is that it is a digital/analog hybrid. Apparently, the chip consists of standard transistors in a ring of artifical neurons and synapses. When impulses hit the neurons, they fire, but they can be regulated by a central inhibitor, blocking an ugly chain effect. The central inhibiting neuron allows control, including filtering of weaker signals to allow stronger ones to come through -- Sarpeshkar compares it to ignoring background noise at a party. It's an interesting concept, at any rate.
---
"Life. Don't talk to me about life."
The real challenge is going to be maintaining quantum states that are stable enough for work to get done. Right now standard computing is either on or off, and if something goes wrong, you just have to set up the computation again.
With quantum computing you may have a lot of work to reset, unless you can find a relatively easy way to generate quantum effects.
JHK
http://www.cascap.org and you'll never know unless you look
There was a conference at Stanford a while back (was mentioned here IIRC) on synthetic intelligence in general; all sorts of fun stuff was tossed out:
http://www.technetcast.co m/tnc_program.html?program_id=82
This quote (from John Holland) is particularly telling: So we're not quite there yet. Hans Moravec participated in the conference as well, and he has a fairly informative essay linked from his site entitled "When will computing hardware match the human brain?":
http://www.transhumanist.com/volum e 1/moravec.htm
:wq
Now, 10% of autistic people have "Rainman" abilities - massive mathematical powers, etc., and apparentrly the current theory is that theses autistics are merely missing the final "step" in calculating things like humans do - the can't get that final estimate which allows us to get by in society easily.
Are really cool machines that are trying to mimic humans ever going to get to stage where they can estimate things, or will they be like Data from Star Trek TNG. Hmmmm....
Acting stupid isn't much fun when there's someone around who knows better
I'm almost a little disappointed to read this coming from MIT, because when I left the University of Manitoba (Canada) a similar project was being given as a thesis project for fourth year students. The prof coordinating it has been doing research on building neural nets with semiconductors instead of software constructs for a while now. Granted, this bit from MIT might be more complex, or introduce new functionality to the neural net (such as the voice recognition system that incorporated time delays in the calculations last year). But it still seems to me that something is only big news if one of the 'big' colleges works on it. Bleah.
When I finish this internship and go back to finish my fourth year, I'll be proud to go to my hometown U. It's obviously keeping up with the rest of the world - the only thing lagging behind is the media's perception.
You know what to do with the HELLO.
You know what to do with the HELLO. ...
Help create an open-source world
http://www.wired.com/news/technology/0,1282,3702 9,00.html
I like how it's called a breakthrough in "neuromorphic" engineering. Doesn't it just become ten times more impressive when it's described in made up technomumbojumbo?
One time I threw a brick at a duck.
Yeah but could you make a...
Oh yeah, they're made to be clustered!
Eh...
I don't agree 100% that the brain makes digital decisions. The article says that we make an either/or decision regarding whether something is there or not. It is a car or it isn't a car. That's rather black and white. If a picture is blurry or if the object is partially hidden, then we could say, "It is almost a car," or, "It might be a car," implying that there is a degree to which something might be a car.
If you run an analog signal through a filter, you can detect if certain frequency is present. This may seem digital, similar to the car case, but actually it can be an analog signal and and analog filter. The results, similar to the car may be that the signal present, but it is not statistically significant above the background noise/interference.
To make a long story short; I still believe that humans make analog thoughts, even if our brain is just one big circuit.
Keeping
The Institute that is doing the research has more information here. I believe the guy doing the actual research has more research here.
Next time you get multiple submissions, try picking the post with more info than the rest instead of attempting to summarize. Especially when you leave out the important links.
--
Gonzo Granzeau
Gonzo Granzeau
"Nothing the god of biomechanics wouldn't let you into heaven for.." -Roy Batty
I've read the original paper in Nature. (I'd post a link, but I only have access via my university's account, and I have no interest in getting that revoked.) This is not exactly a neural network in the classic sense, although it is similar. The standard neural network is specifically designated an artificial network --- it implements a computational model of neurons. These guys are actually attempting to simulate the known electrical behavior of neurons, in the theory that a network composed of elements that truly mimic neurons will be more brain-like.
Now. "Digital and analog." This is not a new discovery. It has long been known that neurons have a specific threshold WRT to incoming signal; if the incoming signal does not meet the threshold, the neuron will not fire. If signal is above threshold, the neuron fires. If signal is really above threshold, the neuron fires repeatedly, encoding the strength of the stimulus as the frequency of the train of pulses. (AFAIK, the circuits described here didn't implement that last behavior.) This is a digital response. The output, however, is a continuous voltage at a particular frequency: an analog signal. (Whoever called this "a digital response to analog criteria" is correct.)
The important thing is that connections between neurons have different weights, and there's often a lot of local feedback. In practice, these feedback loops tend to be tuned so that a given cell will respond only to a fairly specific stimulus (the right light intensity in the right part of your visual field, or facing a certain direction relative to known landmarks, or hearing a sound from a certain direction, for example). These guys have implemented a circuit on silicon that shows the same filtering behavior and also captures the idea that neurons can be "on" or "off".
Yes, this is kind of neat. Yes, it could eventually lead to advances in AI; at the very least, it could provide useful signal filtering for robotic applications. No, it has nothing to do with plugging your Pentium into your parietal lobe or your Mac into your medulla, at least not until our circuit-design ability is so good that we can entirely mimic the black-box behavior of brain areas. (Hint: we don't even entirely understand that behavior for most regions.)
I'm also kind of surprised that this made Nature; there are guys at UPenn who've had working neuromorphic circuits for years now. Then again, it's only in the Letters section, and these new guys worked out some mathematical models for the gain of a neural circuit rather than just trying to copy existing ones.
sig not found
JHK
http://www.cascap.org