I've been mulling over purely probabilistic learners, like the ones described here and here. And I have quantified, why I don't think they will be a complete picture of how we learn.
It is because they ignore what I call probabilistic oracles, although there may be a different name for them. Basically they get all there probabilistic information first hand, they cannot currently incorporate other probabilistic information, from other learners. For example I could not convince them that suns age and change colour over time, without showing them the actual stars changing colour. Whereas we know that stars are likely to change over time as they run out of hydrogen and other lighter atoms, without direct experience.
However if you do allow probabilistic oracles, you get into problems of trust, which oracles do you trust and which don't you.
Hopefully my solution where the trustworthiness of an oracle could be encoded within a program and those programs evaluated to see how well it acts in the real world, might get around those problems. Although these sorts of systems do definately bear more thinking about.