It has come to my attention that current C compilers compile C to assembly and then assembles and links the assembly code. What about a new type of compiler, which just develops an intuition on how to generate the raw binary. I'm thinking of creating the next generation C compiler, it's a trained neural network which compiles by intuition rather than preset assembly instructions. It should probably compile code much faster than current C compilers and we can market it as organic and natural since it's just like a C compiling animal as opposed to the current compiler robots.
Compilers are good at what they do. Do you think a human (huge biological neural network) is faster than a compiler? No.
You can omit the "compile to assembly text" part, but you'll still need to generate some sort of symbolic assembly. Some Lisp compilers actually omit the "compile to assembly text" and just generate internal in-memory assembly which is assembled using its own symbolic assembler. If you actually knew enough about compilers, you'll realize that the assembler is actually one of the least resource intensive parts of a compiler.
Neural networks are nice for certain things, but I don't think this would be the best solution for what you want, although I wouldn't rule it out completly as a strategy for certain optimization tasks.