MIT Researchers Explore How Rats Think

Ant writes "A Nature News article explains that, after running a maze, rats mentally replay their actions backwards." From the article: "As the rats ran along the track, the nerve cells fired in a very specific sequence. This is not surprising, because certain cells in this region are known to be triggered when an animal passes through a particular spot in a space. But the researchers were taken aback by what they saw when the rats were resting. Then, the same brain cells replayed the sequence of electrical firing over and over, but in reverse and speeded up. 'It's absolutely original; no one has ever seen this before at all,' says Edvard Moser, who studies memory at the Norwegian University of Science and Technology in Trondheim."

Comments

back propagation learning algorithm

[S3D]

The result is also of keen interest to those who study artificial intelligence and try to teach computer systems or robots to learn through reward and punishment. Some such systems already work by playing back a sequence of moves so that the computer can identify at which point it made the trial or error.
It's called back propagation learning The algortihm is based on the error propagation backwards from the output nodes to the inner nodes of neural net.

Re:back propagation learning algorithm

[eMago]

It could also mean http://www.answers.com/topic/reinforcement-learnin g?method=22 (Wikipedia itself is currently down). Reinforcement Learning (RL) is about learning from reward - and about finding optimal sequences of action. Especially for learning sequences over time - like the rats - it is THE method of choice. And yes, Artificial Neural Networks (ANNs) are often used for representing the "optimal policy" in RL. The weights in those networks are then altered by the RL process.

You could describe the process in the rats brain as doing a "virtual policy search RL".

Pure Backpropagation for long sequences over time, on the other hand, is quite an intractable problem, because you have to feed so many time-states into the network.

Re:brain == computer

[strider44]

You should perhaps read a book called Gödel Escher Bach by Douglas Hofstadter if you haven't already, which develops and expands that theory. It's *starting* to get a bit old at the moment but it's still absolutely fascinating.

Link to paper (requires Nature access)

[thcooke]

The paper is available at Nature Advance Online Publications - if you have access.

New Science?

[ThePopeLayton]

Sorry boys, But I have been studying Neuroscience for the last 3 years. Every lecture we have had on sleep, particullary REM, has taught that when you are sleeping your neurons will all re-fire in an organized manner. This is when your memories are "consolidated" from semi-long term to long term memories. This is why if you have had a particularly stressful day you can "re-live" that day in your dreams. However it has long been shown that "place cells" or neurons that store spatial location will fire in the direct same sequence in which they fired when the test subject was presented with a spatial puzzle. You can read about this in the book Neuroscience by Kandal.