Slashdot Mirror
Flying By Brain
Garabito writes "Scientists at the University of Florida made a living 'brain' by extracting 25,000 neurons from a rat's brain and culturing them inside a glass dish. Then, the neurons began to extend lines to each other, creating a living neural network between them. The dish had a grid of 60 electrodes connected to a computer running a flight simulator. The scientists were able to train the 'brain' to control the plane in the simulator and to react to conditions of the plane. Are we getting closer to create an artificially made conscious being, or perhaps, a living computer?" AlphaJoe was one of several readers to add a link to Wired's article on the experiment.
In Young Frankenstein, the guy is always saying, That's Frank (short o sound) en steen. In German, the second letter does get the sound, so it's ien, not ein. Unless you're talking about the original. Then it's Frankenstein.
Help! I'm being repressed!
Steve Potter, the former mentor of the UF researcher has a pretty thorough description of it. http://www.neuro.gatech.edu/groups/potter/animat.h tml
I'll bite. No, this doesn't necessarily mean that a rat could be trained to fly a plane. A rat has millions of neurons, but most of them are taken up full-time doing specific things (strangely enough, a lot of that is scent processing). But if you can define goals for the rat, you can probably train it to do a lot of things, including a subset of the plane-flying challenge.
You don't want to think of the neurons as "hardware" exactly, either. The process of building and training a neural network is about replacing the programming component of building a system, not about replacing the hardware. Writing a piece of software to fly a plane by itself is hard work--complicated task, not easily reduced to algorithmic instruction sets. Lots of tiny rule modifications needed to the basic set of "maintain altitude and heading". The trick with neural nets is that you set up the network, and then you train it by trial and error to do the task. It programs itself, essentially.
We can and do build neural net simulations in pure software, which is where most of the research has been done so far. But neural net simulations on computers are VERY computationally expensive and take up a shitload of memory, so there are limits as to how big you can make your simulation and still do anything with it. This is a big problem, because neural nets can potentially do incredibly interesting things (like, say sentience!) if they get big enough--but we don't have computers big enough to model neural nets as complicated as we'd like.
I know the article says that these guys are only using this project to investigate how neurons work in the real world, but the potential applications of this are big. Neural nets using actual neurons, not expensive simulations, could be cheap enough to build and train that they would find commercial uses.
A brain is a neural network. Artificial neural networks were created to simulate them using mathematical models.
"If you think about your brain, and learning and the memory process, I can ask you questions about when you were 5 years old and you can retrieve information. That's a tremendous capacity for memory.
I have to say, I don't remember much from when I was five years old. I remember where I lived and maybe can guesstimate where I spent a specific summer, but most of my knowledge comes from what my parents told me and from little "text" snippets that somehow got stuck in my head (for example, names of cities I visited, etc.)
I can recall some images from the past, but I am not sure whether those are "true" memories or something synthesised by brain to "fill in the blank". This leads me to believe that human memory is rather lossy and large part of what I remember is just a rough approximation of what happened based on a few datapoints that brain actually remembers. Sort of like with people who have a defect in their iris - they still see an image in what's supposed to be a blind spot. This image is synthesised by brain to fill in the gap. Needless to say, occasionaly it turns deadly (especially while driving).
"You mortals are so obtuse." -Q
Well, neurons are living cells... ...and therefore they can reproduce. This is called neurogenesis...and as I understand it can be stimulated by appropriate amounts of neurotrophin and other chemicals.
However, with all animal brains, there comes a point in the creature's development where the death rate is greater than the birth rate. In humans it happens at about three years, if memory serves (heh). If we could manage to find the correct chemical balance to maintain an average cell count indefinately, then perhaps we could devise a dietary supplement that would have the same (or better) effect on humans...
Of course, giving a person a lot of neurons doesn't mean that person will make use of them...
I assume you're referring to this
No its not. Whats commonly referred to as a lobotomy, is to remove or seperate the frontal lobes ( Higher functions ) and not seperate the two hemispheres of the brain.
If a first you don't succeed, your a programmer...
Well I took a course about artificial neural net (not the biological ones like here). But we learn that biologocal neurons learn by repetition and correlation. When a neuron sees a pattern it tries to repeat it. They probably ran the simulator under different conditions. While giving input to the neurons they forced the output signals. (with simple voltages) The neurons learned these output signals. Afterwards, they just had to give the inputs signals and the neural net would automatically give the output signals it got used to.
...) and these are all voltage pulses (caused by chemical reactions and input signal from the computer)in the neural net.
basically the net learns an unlinear function or the inputs. outputi = fi(input1,input2,input3,input4
The second article stated that neurons were given information on the tilt of the airplane:
It seems that this experiment builds on earier research by DeMarse, Wagenaar, Blau, and Potter in 2001 called the the animat. It wondered in a box without goal-specific behavior. However, it also tended to specific patterns and states. That is a very readable article - I highly suggest you read it.
But why did the neurons want to stablize the aircraft? I couldn't find a paper on the aircraft experiment, but a second paper, "Removing some 'A' from AI: Embodied Cultured Networks" (by Bakkum, Shkolnik, Ben-Ary, Gamblen, DeMarse, and Potter, 2004) summarized another experiment where neurons were trained to keep a set distance from an object. The paper is the first article on the same page of publications as the first paper. It seems that the neural network responded nonlinearly - that is, it changed state from one behavior to another one - when the input stimulus frequency was adjusted (correct me if I'm wrong). So by changing the input stimulus frequency, they were able to train the network. I gather that the new experiment simply uses when certain "level = good, nonlevel = bad" stimuli. It's a long way off from Robocop II, but it is a start.
It is impossible to enjoy idling thoroughly unless one has plenty of work to do.
- Jerome Klapka Jerome
What hump?
The quoted dialogue above is a hilarious exchange from an extremely funny movie. They made it in B&W and it still worked in 1974. Today it's quite a cult classic.
I haven't seen a slashdot name of Abby Normal yet and you can always slip brains through slot in door after 5PM.
We did not report LTP because it is NOT LTP. In fact, we are using and effect reported by Eytan, D., Brenner, N., and Marom, S., Selective Adaptation in Networks of Cortical Neurons. Journal of Neuroscience, 2003. 23(28): p. 9349-9356 in which "high" frequency stimulations (once every second) was reported to depress the response of the network while "low" frequency stimulations resulted in an enhanced response. For our system we tied the network's response to the control surfaces, dedicating stimulations on one channel for pitch, and a second for roll control. Each channel is stimulated separately, and the response (PSTH) is recorded. Control movements are proportional to the current error from straight and level by mapping the error (0 to 180 degrees) to the interval 0 to 100 ms of the PSTH and integrating the difference in response before training, to the current or enhanced or depressed levels. The more error, the more the control surface is moved. The networks only gradually control the aircraft since the Marom effect requires over 15 minutes to develop. The two frequencies are then used to adjust these weights (i.e. number of spikes in the PSTH) to produce optimal flight. The neurons/network don't seek optimal flight in the classic sense. Instead, we adjust the weights (using high and low Freq. stims) in the network to produce that result. It is a very simple system and our only interest in it is in terms of those changes within the network and the possibility to extend it to more of the network than just two or three different channels. Hope that helps.. Tom DeMarse