[seek3r00]

Some interesting extracts:

“The instruction set together with the input representations jointly determine the class of algorithms that are learnable by the neural controller, and we see this as a fruitful avenue for future research akin to how current instruction sets shaped microprocessors.”

“The generalization or correctness we aim for is mostly about generalization to instances of arbitrary sizes.”

“[...] computing a = f(s) can be more expensive on current CPUs than executing typical computation employed in the algorithms studied here. We thus hope that this research will motivate future CPUs to have “Neural Logic Units” to implement such functions f fast and efficiently, effectively extending their instruction set, and making such approaches feasible.”