Prof. Mayer Aladjem

Prof. Mayer Aladjem Profile

Associate Professor


Department : School of Electrical and Computer Engineering
Room : 114
בנין המחלקה להנדסת חשמל ומחשבים ע"ש זלוטובסקי - 33
Phone : 972-8-6472409
Email : aladjem@ee.bgu.ac.il
Office Hours :  

Education

  • 1974 Electrical Engineer, M.Sc.
  • Technical University, Sofia, Bulgaria.
  • Department of Automatic and Control Systems Engineering
  • Advisor : Prof.V.T.Kissiov
  • Master's Thesis : "Supervised and unsupervised learning algorithms
  • applied to biological data analysis" (in Bulgarian).
  • 1975 Post-Graduate Qualification "Engineer-mathematician"
  • Technical University, Sofia, Bulgaria.
  • Center of Applied Mathematics
  • Advisor : Prof.V.T.Kissiov
  • Thesis : "Method of feature extraction in two class classification
  • problems" (in Bulgarian).
  • 1980 Ph.D., Technical University, Sofia, Bulgaria
  • Center of Applied Mathematics
  • Advisor : Prof.V.T.Kissiov
  • Ph.D. Dissertation: "Statistical analysis and classification of
  • multidimensional data" (in Bulgarian).

Research Interests

  • Interactive pattern recognitionNeural networks for pattern recognition.Computational statistics and data analysis.Software systems for data analysis.Applications in biology, medicine, character recognition, finger print recognition.

Research Abstract

  • New method for optimization of the discriminant criteria We proposed a new method for identifying two discriminant direc-tions obtained by successive optimization of discriminant crite-ria. It is free to search for discriminant directions oblique to each other and ensures that informative directions already found will not be chosen again at a later stage.The experimental results indicates that the new method provides more efective constraint on the discriminant vectors than does the orthogonal-ity constraint usually applied in statistical pattern recognition. New method for optimization of the Patric-Fisher (PF) dis-criminant criterion We propozed a method for optimization of the PF discriminantcriterion which has several local maxima. For this criterion an iterativeoptimization has to be carried out. We proposed a method of recursiveoptimization for searching the directions corresponding to several largelocal maxima of PF criterion. We carried out experiments with syntheticand real data sets which show that our method was more effective than themethods for initialization of the local optimizer of the PF criterion usedin the past. Neural networks for discriminant analysis We proposed a neural network(NN) implementation of our previous method for optimization of the PF discriminant criterion. We used an auto-associative network havingnon-linear activation functions instead of a linear transformation performed in our previous method. This makes it possible to apply the NN implementa-tion for the recursive training of a multi-layer networks for classification. We have compared our recursive training and conventional training with random initialization of the weights using synthetic data set and an OCR problem. The results ob-tained confirm the efficacy of our recursive training.

Additional links