Exploitable results High-performance classification based on neural networks with fast training The main topic of this project was connected with the development, investigation, and application of high performance neural network classifiers. The classifiers are feedforward multilayer perceptrons with single layer of trainable connections, providing fast learning and global convergence. Therefore the main peculiarity of the developed classifiers is the transformation of the input space performed by the input untrainable (unmodifiable) layers that allows arbitrary complex class boundaries to be formed. In the Random Threshold Classifiers (RTC), an intermediate binary neuron layer was introduced which analyzing whether a point of the input space belongs to a set of overlapping hyper-rectangle receptive field randomly located in the input space, each corresponding to certain binary neuron. The number and the size of those hyper-rectangles was chosen so as to provide a small fraction of active neurons of intermediate layer for each input. Since only comparison operations are necessary to determine the activation of the binary neurons, such a transformation of the input space is computationally efficient. This approach has been refined in a new Random Subspace Classifier (RSC) by using only a small subset of input dimensions in each receptive hyper-rectangle. This approach is 5 - 25 times faster than RTC with similar error rate, and is especially well-suited for high-dimensional input spaces. The versions of RSC for binary input space (in particular, 2D retina for input of binary images) have been also elaborated. To test the recognition rate and other performance characteristics of the developed classifiers and for comparison with existing classification schemes, a generator of artificial datasets with controlled characteristics has been elaborated. The testing results showed RTC/RSC classification rate and speed at the level of best existing classifiers, and reduced training time. The possibilities of an efficient parallelisation of developed classifiers for very large pattern recognition problems have been also considered in the course of the project. The RTC/RSC classifiers have been implemented in the frames of Pattern Recognition Toolkit prototype developed in the project. They have been also applied to a number of real-world problems, including image recognition (OCR of printed texts, texture recognition, writer identification, handprinted digit and character and handwritten word recognition), acoustical recognition (acoustical diagnostics of diesel engine, speaker identification), and some other problems (identification radio wave modulation, control). Recognition of handprinted characters, image texture recognition, and speaker identification were identified as promising applications for further work. Genetic algorithms have been investigated to optimize the structure of untrainable layers of developed classifiers (that can be considered as secondary feature extraction layers) in order to increase the recognition rate. This work has lead to the conclusion that at present more reliable and less computationally demanding improvement of the recognition rate can be reached using "hybrid" classification schemes proposed in the coarse of the project. Such schemes combine the classifiers with different types of input features to improve the recognition rate. Searching for OpenAIRE data... There was an error trying to search data from OpenAIRE No results available
