We hope to have taken one step further in understanding the relationship between programming languages and ARNNs. We started presenting a theoretical framework of a very simple programming language. Next, we described a systematic way of obtaining ARNNs from programs written in the PRF language and concluded with the presentation of a compiling tool. Although the ideas here presented may seem oversimplified when facing real-world applications, the problem itself (can some type of ARNNs perform exact symbolic computation?) has an affirmative answer. Moreover, there is an automatic translation from the mathematical description of any Turing computable function into a Neural Network able to compute that function. This establishes a computational equivalence of this simple model of ARNNs and Partial Recursive Functions (itself an Universal model of computation), which is modular-oriented (making complexity management of large programs easier to deal with). Adapting them to other more elaborate frameworks like those of high-level languages is straightforward, as in [5]. Finally, we also defend that the search for a more suited language to the architecture of ARNNs can give us fresh views on efficient approaches for systematic construction of ARNNs.
展开▼
