Evaluation of a Resource Allocating Network with Long Term Memory Using GPU

Abstract

Incremental learning has recently received broad attention in many applications of pattern recognition and data mining. With many typical incremental learning situations in the real world where a fast response to changing data is necessary, developing a parallel implementation (in fast processing units) will give great impact to many applications. Current research on incremental learning methods employs a modified version of a resource allocating network (RAN) which is one variation of a radial basis function network (RBFN). This paper evaluates the impact of a Graphics Processing Units (GPU) based implementation of a RAN network incorporating Long Term Memory (LTM) [4]. The incremental learning algorithm is compared with the batch RBF approach in terms of accuracy and computational cost, both in sequential and GPU implementations. The UCI machine learning benchmark datasets and a real world problem of multimedia forgery detection were considered in the experiments. The preliminary evaluation shows that although the creation of the model is faster with the RBF algorithm, the RAN-LTM can be useful in environments with the need of fast changing models and high-dimensional data.