biologie, implantaat, ant on the beach, neuropsychologie, robots, brein
Rat-brain flies plane
Old news, but new to me, a piece of rat-tissue was trained to control a flight simulator!
What I understand of the web-article is that they used biological neurons to create a network of cells in the same manner as they do in software ‘neural network’ systems. Artificial (software) neural networks work on some kind of local learning principle, for example some kind of error-feedback learning. Imagine you give a network some kind of input, and it will generate a (random) output. You then tell the network whether it did right or wrong and the network will adjust the strenghts of the connections between the neurons a bit, where ’succesful’ connections become stronger and the ‘wrong’ connections become less strong. Step by step the network will learn to associate each possible input with the correct output. In cases of a flight simulator the situation is a bit different but the principles stay the same. Here the network works mostly like cybernetic control system where the succesvalue of the output is continuously calculated and continuously fed back into the system as a constraint, driving the system towards the right kind of behavior. Reinforcement learning might be what these researchers have used here but I don’t know about the details of the study.
So now they used biological neurons instead of software neurons to do the same trick. BUT, that of course doesn’t mean that this is the way these neurons work in real brains. The researchers *used* the neurons to create a specific part of a system that is designed by researchers (an artificial neural network). It’s still artificial. For example, the error-feedback procedure used by the researchers might not at all resemble the way real brains correct their activation on basis of error feedback (if they do so at all). The big problem of understanding real brains is: WHO determines what is ‘right’ or ‘wrong’? Who puts the ‘value’ on the feedback signals that are to be used to change the network connections in the network? In other words, how does a real brain know it is ‘in a flight simulation game’ and how does it know ‘what is good and bad’ in a flight simulation game? If nobody tells the brain if the direction it’s going is good or bad, then it cannot learn anything. But we do learn, and nobody ‘tells’ us. In the end it has something to do with survival, with ‘keeping yourself satisfied’. At the moment you do something that gives no satisfaction, then the brain will probably receive this as an error-signal. But then the question becomes: how does the body monitor ’satisfaction’? For food intake or body damage I can image how this would work but for “flight simulation games”????
FURTHER THOUGHT Of course not only our body tells us something about ‘good’ and ‘bad’ but also the physical possibilities for behavior in the real world. If you crash your plane, then obviously you did something wrong. The big thing of being ‘intelligent’ however is that you can learn to fly a plane without crashing even once (since crashing will also end you life so you wouldn’t get to much learning experience out of that anyway). So how can we learn from the actual physical feedback we get from the world (e.g. the way the steering wheel feels in our hands and the way the visual flow changes in our eyes whilst we fly the plane) and then project that sensory feedback onto the succesvalue of a task that we do not get *real* physical feedback from, that is, how do we use the ’safe’ information (visual flow, sense of gravity) and use that to learn how to correct our errors *before* we actually crash the plane? In the end, this will probably mean that representation has to kick in at some point. So for example I see how the horizon tips over to the left, and somehow my brain has to take that feedback and make it mean something else than just ‘the horizon is tipping over to the left’. My brain has to see it tipping and then ‘conclude’ in some way that this *means* that my plane is ‘going to crash soon if I do not act quickly’. Gibsonian, ecological psychology would hold that representation is not necessary here, I don’t know who’s right. Still even in this case I still miss the ‘good’ and ‘bad’ valorisation: if I see the horizon tipping to the left (and feel gravity changing at the same time), how do I know this means I am doing something wrong? I still won’t know until I actually crash, right?
Popularity: 18% [?]
21 Mar 2011 admin 0 comments