Slashdot Mirror
In Motor Learning, New Brain Connections Form Rapidly
Science Daily has a report on research demonstrating directly that new connections begin to form between brain cells almost immediately as animals learn a new task. A team lead by researchers at UC Santa Cruz performed "...detailed observations of the rewiring processes that take place in the brain during motor learning. The researchers studied mice as they were trained to reach through a slot to get a seed. They observed rapid growth of... synapses between nerve cells in the motor cortex... The study used mice that had been genetically altered to make a fluorescent protein within certain neurons in the brain. The researchers were then able to use a special microscopy technique (two-photon microscopy) to obtain clear images of those neurons near the surface of the brain. The noninvasive imaging technique enabled them to view changes in individual brain cells of the mice before, during, and after the mice were trained in the seed-reaching task."
So, what connections are formed when one learns to drive a motor vehicle?
Thats how long it seems to take between starting to learn something and getting it properly. Learning a new, totally different programming language would take that long for me. But the transition, when it comes is very fast.
http://michaelsmith.id.au
I think you mean artificial neural networks (ANNs). Yes, they are supposed to be similar to biological brains but the devil is in the details. There is some question, at least among computer scientists, about _exactly_ how the biobrain does it. Gerald Edelman put forward some interesting ideas in the book _Neural_Darwinism_. Ken Stanley has been working on something called NEAT for many years, building on it with Compositional Pattern Producing Networks.
Refining our observations of how nature does it may help produce better artificial neural networks, among other things...
Coarse muscle control comes from your motor cortex, fine control and feedback loop for motor movement through the posterior lobe of your cerebellum. Driving a motor vehicle requires in fact a lot of motor coordination, most people can learn how to. Note, however, that making an intelligent joke in the first post requires a much more complicated mechanism.
My other signature is a car
... but reconfiguring the body from bones up (to compensate for aquired deficits regarding posture, e.g. by continued abuse while sitting in front of a screen) takes much longer. At least, this is my experience.
CC.
TaijiQuan (Huang, 5 loosenings)
Does anyone have a car analogy to help with this one?
Summation 2
I get it. the brain is big enough for me, my alternate personalities, my paranoid delusions, my subconscious and the useful part dealing with the heartlungsstomach thingies.
I thought various experiences just put information in my memory. But if experiences can actually rewire the brain, where does that lead me? is "I" rewirable too?
Thanks for nothing. now my day will be spent on purposeless pondering of the nature of me.
new sig
Most neural networks have fixed layout of "synapses" and operate only on changing neuron thresholds.
Thing is the thresholds are a temporary, easily forgettable kind of memory. The synapses once grown stay there for the rest of your life (unless you kill them, say, with alcohol).
This explains why if you learn riding a bicycle once, you need up to 15 minutes to recall it even if you haven't been riding it for some decades. Same goes for swimming.
45 5F E1 04 22 CA 29 C4 93 3F 95 05 2B 79 2A B2
Current student of Ken Stanley in Neuroevolution and Generative Development this semester.
Parent is correct: The devil is in the details. A Neural Network is a _model_ of what actually happens in the brain. It is an approximation. There are a number of things that the model does not account for, such as the growth of new connections (somewhat accounted for in the GA-NN NEAT), and the exponential response nature (accounted for in CRTNN networks).
CPPNs are a model to account for generative development, rather than Neural Networks. The hope is to get a full system without going through the actual step-by-step constructive development of it. To this end, it is successful.
You can find more information about the subject, or implement your own CPPN network here: http://www.cs.ucf.edu/~kstanley/neat.html
The article presents a good argument that the ANN model is at least partially incorrect on its approximation of brain development. ANNs do not add connections after the topology is created. This could provide interesting new developments to the AI crowd.
I just learned kung-fu.
It'll be funny if it turns out that the neurons are actually quite smart and their only problem is they don't have much in the way of arms, legs etc to prove it :).
In the last decade or so, research has finally shown that new synapse formation and pruning continue to occur throughout life, not simply during the critical period in our early years. (the main article here is one such example). So, while the statement about re-learning a bicycle is likely correct, to say that most of our neural networks are fixed and operate only on changing synapse thresholds is questionable.
"Owning a computer is like having your very own TV -- with a built in radio!" - Ed Helms
Nope, not what I meant. Most of -artificial- neural networks are fixed. The mathematical/software constructs usually display a fixed, organized architecture with little or no ability to create new connections from scratch. The self-constructing neural networks are a small margin of the science.
In most cases you set a number of neurons, connect them in a specific pattern, then save the construction and run the network: feed the input, analyze the output and adjust weights so that some connections are more active and some are less.
Meanwhile biological brain grows new synapses - which is different from creating new neural paths through existing synapses.
45 5F E1 04 22 CA 29 C4 93 3F 95 05 2B 79 2A B2
Comment removed based on user account deletion
Dogma has been that the brain does two kinds of learning - short term and long term. Short term learning happens within a few (or even a single) depolarization, lasts for a few hours, and is perfect for learning that the path is wet today. Long term learning has been seen as a separate, but related process, where repeated neural activity triggers new protein synthesis, and that synthesis results in new synapse formation. That process is thought to require repetition over minutes to hours, results in learning that lasts for days to weeks, and is well suited for learning that this path goes to grandma's house.
To find synaptic remodeling after a single training trial would require some revamping of that model. It seems reasonable enough, especially in more complex brains. A lot of what we know about the biochemistry of learning comes from invertebrates with fewer than 1e5 neurons. Even a mouse brain has ~1e8 neurons, which means there are a lot more opportunities for reinforcing signals, internal repetition, and god knows what else that might accelerate the long-term learning process we see in invertebrates.
>>There are a number of things that the model does not account for, such as the growth of new connections (somewhat accounted for in the GA-NN NEAT), and the exponential response nature (accounted for in CRTNN networks).
It has long been established that physical movement is a major neurotropic factor. I can dig up the reference if you want.
It's always amusing when Slashdot touts something as amazing when it's been known for a long time.
Having hiked the Flume, I'd never go _DOWN_ the slide trail. As an avid hiker, the synapses in my brain do everything possible to avoid steep descents.
Comment removed based on user account deletion