Second draft... wrote friday 6th
I'm sitting in la maison rouge near place de la Bastille, attending the presentation during Intensive Science, the open days for the 10 years of Sony CSL Paris.
Currently Frédéric Kaplan is speaking about his effort and research involving creating a playground for AIBOs. More than the results of the research itself, that I find charming but probably a little bit out of my reach, what I found very useful was the breakdown of the cognitive process of learning that lead the researchers to estabilish a model for curiousity driven behaviour.
Basically they found out that it was really difficult to teach concept to AIBOs because they lacked attention.
In order to get attention, they needed a know-how, that is know what to do with a specific feature of environment; to understand this concept, try to think of a dining room from the perspective of a man and a fly: the man will see all kind of objects (table, chairs, plates, glasses, couch, bookshelf, and so on), while the fly will "perceive" just the lightbulb and probably the plates and glasses, the rest will be just environmental noise.
But where the know-how origins? It comes from motivation.
So the next step is to give our learning subjects (AIBOs in this case) a motivation to learn. The problem is: how to model it?
The key factor chosen was error predictability: different behaviours will choose "what to do next" based on the fact that the task will have a high, moderate or low chance of failure, from daredevils to sheep-mode that is.
To cut it short, they found out that neither choosing the high risk nor the low risk was good for the learning of the AIBOs (in the first case it would be stuck with some very unpredictable task, in the other will just shutdown because it's the safest choice), but the best results come from studying a function of the derivate of risk prediction: engage in those tasks that are challenging at first, but will then become easier and easier.
Which is something we instinctively do constantly, I know... but it's nice to see it shown as a math model :)
Leave a comment