Gresham College lecture on deep learning

It was great fun to remind myself of the history of deep learning. Looking at the lecture I wonder I come across as to sceptical about deep learning. It's true that the effectiveness of deep learning is unmatched by other technologies but it's also true that people are having their papers rejected if they don't use deep learning - and there may be very good reasons why academic papers don't use deep learning. The principal objection, and this is mentioned in the lecture, is that deep networks, in fact shallow networks, give one little indication of how the inference was performed. So if your research has a scientific emphasis, rather than an engineering one, its natural to want to explain the why and how rather than to revel in some fantastic result. Of course there is a place for both -- without fantastic benchmark results how do we know if anything is worth doing?

I'll make a note to return to this and other topics in AI in later blog posts (I hope).