Russell and Norvig [0]. The best introduction IMO to "real" AI, as opposed to the ML stuff that currently passes for AI.
Until you can answer these questions, you're not doing science. You're just throwing data at Tensorflow or Keras until you guess it's good enough. And then someone dies in a self-driving car accident because the network didn't learn that a transverse truck or a pedestrian was an obstacle.
ML does impressive stuff but as several recent HN articles have pointed out it's basically just curve fitting. And voodoo curve fitting at that.
>"hit something of a brick wall decades ago"
Is true. Why do you disagree? What improvements did "old-school, mostly symbolic AI" bring to the current field of research?
Sure, ML has failures - but those failures are in applications and fields where old school symbolic AI can't even reasonably be applied to. We have to start somewhere and just using symbolic AI is far behind in terms of the requirements we have currently.
>"How many layers do you need and why? How many training cases do you need and why? What has the network learned and how do you know that? What important things has the network not learned? When will it fail?"
A lot of these issues have been addressed in many recent papers. A lot of these papers have been solely focused on understandable/explainable machine learning which is an overarching topic that covers all your questions.
>"Until you can answer these questions, you're not doing science."
So, you are essentially saying a large part of CS academia is not doing "science". I'm not sure what kind of "science" you have done to make such comments. But I'm pretty sure there are plenty of researchers out there who are far more of an expert than you are in this field would wholly disagree with you.
Neural nets are also showing a lot of promise for symbolic tasks like automated theorem proving. So I would predict that in time they eventually become a standard technique in what would otherwise be the domain of older techniques.