Supporting Discovery Learning in Building Neural Network Models

Abstract

In this paper, we propose a framework based on a hybrid approach to support learning neural networks within an interactive simulation-based learning environment [1], allowing learners to build their neural network simulators by blending the theory with praxis. To diagnose a neural model made by a learner, we construct a script file during the learner’s manipulations of objects using a set of inference rules to help determine the network’s topology and initial weight values. To this end, we embedded in the system a virtual assistant (VA), which also contributes in the educational stage of neural networks usage. The VA uses the knowledge-based neural network (KBNN) algorithm [2] to translate the script file into a set of nodes represented as an AND/OR dependency tree. From the tree we choose a proper index and examine whether the architecture and the corresponding parameters can be approximated in the knowledge-based neural network (KBNN) space. The VA examines if there is any missing information or wrong con ception and points it out to the learner showing him/her where the error or misconception might be. The system has an object-oriented architecture in which an adaptive user interface connected to the learner’s skills has the role of a motivator in the learning stage. That is, it allows visualizing neural models as concrete neural objects. In fact, the most of the existing neural networks models have some common components [3], the idea is to implement those simple components and use them to build different a nd even very complicated systems.