Inductive Inference of Tree Automata by Recursive Neural Networks

Recurrent neural networks are powerful learning machines capable of processing sequences. A recent extension of these machines can conveniently be used to process also general data structures like trees and graphs, which opens the doors to a number of new very interesting applications previously unexplored. In this paper, we show that when the problem of learning is restricted to purely symbolic data structures, like trees, the continuous representation developed after learning can also be given a symbolic interpretation. In particular, we show that a proper quantization of the neuron activation trajectory makes it possible to induce tree automata. We present preliminary experiments for small-size problems that, however, are very promising, especially when considering that this methodology is very robust with respect to accidental or malicious corruption of the learning set.