I don’t know any detriment, so here is my denial:
- The “rules” that come from neural networks are not simply “ IF A THEN B”. There is at least some arithmetic processing, at least a linear combination of input values. And that’s the simple neural nets that I learned long ago and I’m sure there is more involved now.
- It’s very significant that the computer learns these rules on it’s own and could have huge resources, which can include huge quantities of data processing power, in doing so.
- Finally, what are you comparing with? One possibility is humans. Well, what do humans have that cannot be emulated with deep neural nets given time and resources (yes, I agree it will take some time)? The other possibility is that you are comparing with computational completeness. Here’s the only part that I am unsure of and where you might have a point worth investigating. I don’t know whether and how deep neural nets can be compared with computational completeness (going back to Turing). I’m don’t even remember exactly how to define computational completeness. But I can understand how the generality of computer algorithms might seem broader then deep neural nets in some way. It’s a good question to research.