"Tailored" Neural Networks to Improve Image Classification Lerner, B., Guterman, H., Dinstein, I. and Romem, Y.* Department of Electrical and Computer Engineering Ben-Gurion University of the Negev Beer-Sheva, Israel 84105 * The Institute of Medical Genetics, Soroka Medical Center Beer-Sheva, Israel 84105 Abstract The concept of "tailored" neural network is inspired by the concept of grouping in the visual cortex of the mammalian brain. This biological animated concept was implemented to develop "tailored" neural networks for image classification improvement. Each "tailored" network was specialized to classify a different class of vectors. This was done by employing separate training and using specific features in each class. Image classification improvement was tested by the chromosome classification application. For chromosome classification, the probability of correct classification using the "tailored" networks was 2.5% higher than the probability achieved by a conventional neural network (97.6% versus 95.1%). This improvement was found to be higher when lower quality features were employed. It is expected that the improvement will increase whenever the image classification task will become more and more complicated. 1. Introduction Image classification using multilayer perceptron (MLP) neural networks has become widespread in the computer vision and neural networks communities. The neural network classifier has the advantage of being fast (highly parallel), easily trainable and capable of creating arbitrary partitions of feature space. However, image classification using an MLP depends on a series of various procedures generally held according to practical considerations. In most vision applications these stages precede the classification itself and are motivated by a mathematical analysis and/or engineering concepts. Even the MLP classifier itself, when applied to a complicated classification task fails very often to correctly classify the input data. The mammalian visual cortex seems not to suffer from this kind of problems. It simply does not function as our classical image classifiers do. Image projected from the retina onto the visual cortex parallelly spread among a series of cell clusters, each of which performs its own special analysis and synthesis. Each retinal area is analyzed over and over again, column after column, and again in neighboring cortical regions, with respect to a number of different variables such as position, orientation and color [2]. From a large series of experiments, it became apparent that in area 17 simple and complex neurons with similar receptive field axis orientation are neatly stacked on top of each other in discrete columns. Separate columns exist for each axis orientation. Other functional variables are also grouped in columnar aggregates of cells. In cortical areas of the monkey beyond area 17 of the visual cortex, there exist columns of cells with well-defined color sensitivity and other columns in which the direction of movement of the visual stimulus is important. Cortical structure and functional organization go hand in hand [2]. ____________________________________________________________________________________________________________ # This work was supported in part by the Paul Ivanier Center for Robotics and Production Management, Ben-Gurion University, Beer-Sheva, Israel.