Neural networks for processing data structures

Abstract

The possibility to represent and to process structures in a neural network greatly increases the computational capability of neural networks. This new capability, besides to provide a new tool for the classification of structures, can also be exploited to integrate neural networks and symbolic systems in a hybrid system. In fact, structures generated by a symbolic module can be evaluated by this type of networks and their evaluation can be used to modify the behavior of the symbolic module. An instance of this integration scheme is given, for example, by learning heuristics for automated deduction systems. Goller reported very successful results in using a Back-propagation Through Structure network within the SETHEO theorem prover [8]. On the other side, it is not difficult to figure out, in analogy with finite state automata extraction from recurrent networks, how to extract tree automata from a neural network for structures. This would allow the above scheme to work on the other side around, with a neural module which is driven by a symbolic subsystem.The computational results presented in this tutorial, however, need to be extended to DOAGs in order to fully characterize the kind of symbolic computations which can be performed by recursive neural network. Nevertheless it must be pointed out that, due to its numerical nature, recursive neural networks perform a kind of computation which is not possible to easily reproduce by a symbolic system. The impact of this new style of computation should find very useful application in domains where structure and numerical values are both important aspects of the computational problem, such as in the prediction of the biological activity of drugs in Chemistry.KeywordsRecurrent Neural NetworkFinite State MachineHide UnitRecurrent NetworkTree AutomatonThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.