Linguagens de domínio específico e sensores baseados em modelos biológicos de computação

Abstract

A Domain Specific Language is a specification language dedicated to a particular domain, representation technique, or solution searching method. On the other hand, a general-purpose programming language is a language designed with the goal of emulating Lambda Calculus or Turing Machine. Since general-purpose languages must accept any algorithm that can be executed by a Turing Machine, they do not contain a knowledge base of a specific domain, which makes them difficult to master by a professional who is not a specialist in Computer Science. Many fields of science and technology have well advanced domain specific languages: LATEX and XML for text processing, SQL for data base management, Matlab for engineering, etc. Designers of neural network applications do not have good domain specific languages. The main reason for this situation is that they are considered computer scientists, and supposed to know general-purpose languages and even low level languages close to Turing machines (languages for microprogramming). Since we know that this has no base on reality, the goal of this work is to develop methods of creating domain specific languages rooted on Lisp macros and on functional languages (Clean and Haskell) for prototypes. We believe that Lisp is easy enough to be mastered even by people who has difficulty with formal methods and mathematics. Therefore, with well designed extensions, and a rapid training in programming methods, an engineer can use Lisp as a powerful tool of productivity. Lisp is a traditional tool for creating both embedded languages and domain specific languages. Therefore, Lisp offers all tools necessary to design, test and deploy domain specific languages. Nevertheless, neural networks have features that require special care. For instance, they need to have biological feasibility. In order to reach code quality close to human designed nets, while preserving biological feasibility, this work will use advanced Artificial Intelligence methods and guided automatic theorem provers similar to the one used in HOL4 or Isabelle. Our system has a simpler goal than general theorem provers, since it will focus in a few problems related to neural networks. One of these problems is the designing of devices for data acquisition. This thesis differs from similar works in one point that may cause controversy: Besides the contributions of the author to the state of the art, it discusses elementary aspects of the technologies involved. We decided to do this in order to bring the ideas that we are exposing to the reach of all readers. What are these ideas? The first contribution is to bring software engineering technology to biological modeling. We are not aware of any researcher who attempted this goal before. A second aspect of this work can be resumed as an effort to create a language to describe combinations of neuron models into complex networks. All programs discussed here were designed to avoid obsolescence. The author achieved this goal by writing prototypes in ¸-calculus, a formal system that can be executed in modern computers.