Introduction

The application INNE (Interactive Neural Network Environment) is a graphical environment to design, simulate and analyse the behaviour of neural networks. It has been developed in the European project CoLoS (Conceptual Learning of Science) [CoLoS] [Här94], which has involved several European Universities and research centers since 1988 when it started under the auspicies of Hewlett-Packard. His objective is to support concept learning and deeper understanding in the field of science and technology making optimal use of the didactical potential of modern communication technologies.

The aim of the application is to provide a rich and flexible tool for learning neural network modelling. This goal has influenced several design choices and in particular the emphasys that has been given on showing the dynamical processes that occur in the net during the learning phase and the computation.

The neural models implemented are:

Boltzmann machines and Hopfield networks, i.e. combinatoric optimization models with networks that do not change dynamically their structure [AHS85].
Back-error propagation networks, supervised models where learning takes place under supervision, that is confronting the output with the correct answer [Rum86].
Hebbian networks, unsupervised models where learning takes place with no feed-back from the environment and with redundancy in the input data [Oja92];
Simple competitive networks, unsupervised competitive models where single neurons compete for being the one to fire [Her91];
Kohonen networks, unsupervised competitive models where single neurons compete for being the one to fire, as in simple competitive models, and also perform feature mapping [Koh88].

Given its didactical purposes, the application provides a number of pre-defined examples in order to teach the main concepts concerning a chosen algorithm while easing the process of gaining experience on the net architecture (i.e. acquiring the ability of defining the network structure and parameters). In fact, the ability of designing neural architectures to solve problems is acquired only by experience, since the demonstration of convergence is based on stochastic approximation techniques. On the other hand the cardinality of the training set is usually very limited and the choice of the parameters that regulate the network behavoiur must be found heuristically in several trials. Furthermore it does not exist yet a general theory associating the number of learning steps to the approximation reached and this makes it difficult for beginners to start solving problems by adotping the neural paradigm.

In order to realise a range of examples easy to understand it has been necessary to enrich INNE with some auxiliary tools, for instance a generator of gaussian distributions and a generator of uniform distribution inside polygons. These tools are used to generate strong visual flavoured examples so that one can ``see'' what is going on during the computation.

Moreover to make examples more concrete and allow students to fully appreciate the implemented models, the environment provides a range of problems solved with neural networks with which students can experiment. For instance we illustrate the principal components analysis with the problem of image compression, the clustering analysis with the problem of image segmentation, the problem of topological mapping with an example of approximation of bi-dimensional figures.

INNE architecture

INNE architecture INNE: a Neural Network INNE: a Neural Network