Closed pevnak closed 1 month ago
Karel's suggestion is to go through all rules and see, if we can separate all left hand sides of rules, because If yes, then we can recognize all of them.
We need to encode numbers.
:(a + b)
should be different to :(b + a)
that they are the same, because we do not encode the name of variable. Karel suggest to renamve variables using De Bruijn encoding (indices), which just renames from left to right. Project numbers through hyperbolic tanh.
Find 2 states (expressions) which are not recognizable by a current neural network.
Create a random neural network
f
and you have two expression ex1, ex2if
f(ex1) == f(ex2)
then very likely ex1 and ex2 are unrecognizable by your NN architecturePossibly try position encoding