Rescorla & Wagner (1972)
Would it be possible to talk a little more about how exactly associative strength (or what you call 'weights' in the lecture) is calculated? In other words, this seems a lot more straightforward when one is testing a model, but when one is trying to test these predictions in humans or even animals, is there a straightforward way to assign or to precisely estimate a value of associative strength to a given CS-US pairing in a specific case?
Reading about blocking went against my intuitions. Despite some empirical evidence is provided, and despite the fact that, as pointed out by the authors, this is perfectly accounted by their model of Pavlovian conditioning, it is surprising to think that no learning occurs when a new CS is added to the pairing. I find this particularly surprising when thinking about instances in which one might face danger (such in the experiment with rats) since in these contexts in particular, it is highly inefficient and potentially life-threatening not to learn the new association. Perhaps I missed this but is it possible that they are learning that the two new CSs -together- are associated to the US, hence why when tested on the new CS alone they don't show a response?
Lake et al. (2017)
The authors of the paper talk about the role of neuroscience in inspiring AI. The importance of neuro-scientific evidence in supporting theories of brain function is not talked much. If a computational model can predict human behaviour, can we ever reliably say that it has anything to do with how the brain works unless we are able to find evidence in the brain?