A Fast Self-Organizing Map Algorithm for Handwritten Digit Recognition

This paper presents a fast version of the self-organizing map (SOM) algorithm, which simplifies the weight distance calculation, the learning rate function and the neighborhood function by removing complex computations. Simplification accelerates the trai

  • PDF / 1,801,717 Bytes
  • 7 Pages / 439.37 x 666.142 pts Page_size
  • 70 Downloads / 237 Views

DOWNLOAD

REPORT


Abstract This paper presents a fast version of the self-organizing map (SOM) algorithm, which simplifies the weight distance calculation, the learning rate function and the neighborhood function by removing complex computations. Simplification accelerates the training process in software simulation and is applied in the field of handwritten digit recognition. According to the evaluation results of the software prototype, a 15–20 % speed-up in the runtime is obtained compared with the conventional SOM. Furthermore, the fast SOM accelerator can recognize over 81 % of handwritten digit test samples correctly, which is slightly worse than the conventional SOM, but much better than other simplified SOM methods.





Keywords Neural network Self-organizing map Handwritten digit recognition Simplification



1 Introduction The self-organizing map (SOM) also called Kohonen neural network is a competitive learning artificial neural network proposed by Kohonen in 1982 [1]. It is an unsupervised learning method which has both visualization and clustering properties by discovering the topological structure hidden in the data sets. Essentially the goal of a self-organizing map is to map continuous high-dimensional data onto a discrete low (typically one- or two-) dimensional feature map. Y. Wang (&)  Y. Pan  X. Yan Institute of VLSI Design, Zhejiang University, Hangzhou, People’s Republic of China e-mail: [email protected] A. Peyls  L. Claesen EDM, Hasselt University, Diepenbeek, Belgium

J. J. (Jong Hyuk) Park et al. (eds.), Multimedia and Ubiquitous Engineering, Lecture Notes in Electrical Engineering 240, DOI: 10.1007/978-94-007-6738-6_23, Ó Springer Science+Business Media Dordrecht(Outside the USA) 2013

177

178

Y. Wang et al.

As a clustering algorithm, the SOM has been applied widely in various fields including pattern recognition, defect inspection and as a data-mining tool to perform classification of high-dimensional data [2, 3]. Research on improving the performance of the SOM has been going on for decades. One of the key issues to overcome is the low speed learning process while obtaining a well trained map. A SOM is well trained if clustering is achieved in a short time and, at the same time, it creates a projection of data into the map strongly related to the distribution of data in the input space. One of the main reasons for this continued research effort is that the amount of data which is to be analyzed can be huge, for instance thousands of high-dimensional image vectors. The simulation of extensive networks with thousands of neurons, each with high-dimensional weights takes relatively much time on state of the art general purpose computers. To solve this problem, this paper presents a fast version of the SOM algorithm and software simulation proves that the SOM has been accelerated to some extent. The remainder of this paper is organized as follows: Sect. 2 gives a brief overview of related works. Next, Sect. 3 presents the conventional self-organizing map. Section 4 presents our proposed fast self-