eec

E E C !

What the network does not have is any sense of relevance of what it is doing. And this is reflected in the learning mechanism. The network learned on a recurrent version of backpropagation. The basic thing to get from backpropagation is that the behavior of the network is tested against an ideal behavior. But it is you, modeler, who decides what is ideal. In the real-world there is no divine instance telling you what is good or bad. Some way or another biological systems learn by reference to what is relevant for them. In some way or another our behavioral patterns and the mechanisms that change these patterns are embodied in the biological needs and embedded in the environmental demands of our species. And ofcourse I already knew this, because I participated in this reading group on EEC!, that is, Embodied embedded cognition. So, luckily for me, I got a chance to merge these ideas into my thesis. (My supervisor being Pim also helped).

Neural networks up to date have two main problems, and this network was no exception.

Neural networks (especially so called connectionist or PDP networks) learn from examples that are carefully chosen by the modeler and presented randomly one after another in a special training session. This means that the interaction I talked about earlier is kind of artificial: the behavior that the network shows after receiving some input is not influencing which next input the network gets to see. So any temporal correlation real organisms might detect in a sequence of interactions with an environment cannot be exploited by these networks.
Neural networks judge the succes of their behavior (and hence the direction of development) by reference to ideal patterns carefully chosen by the modeler. No reference is made to whether the behavior is really succesful for the agent. (Whether your behavior ultimately let's you eat, drink, avoid predators, mate and live happily ever after)

Thus, in short, I proposed the following:

Let every "output" of the network have a real causal effect on the environmental variables
… so that every "input" to the network is determined in part by the previous output
… so that we have real interaction instead of random input-output mappings.
Get rid of this ideal behavior variable
Let the network judge its own interaction by reference to a biologically plausible variable
… such as, for instance, a bodily parameter that rises and falls
… in reaction to the specific environmental condition one is in.

This way, no external observer is judging the relevance of the performance (the error).

Hopefully, such a mechanism will lead to a situation where, in case of my network, the 4 relevant patterns are weighted more heavily than their mirrorpoints in activation space, so that you will really do the right thing when the lion is alive and the cage door is open. This is however still to be implemented and tested in the model.

THAT'S ALL FOLKS, or was it . . .

back

top