Research Interests

• Neural Networks,
• Artificial intelligence, and
• Genetic algorithms
• Cryptography

Neural Networks

Reading for MPhil at University of Ruhuna on Statistical Mechanics of Cross Model Neural Plasticity (supervisor: Dr. J.R. Wedagedara)
• project proposal (PDF)
• Poster (PDF) [Educational Exhibition at University of Kelaniya, May 2009]
• Presentation (PDF) [as a requirement for the Staff Development Program at University of Kelaniya, July 2009]


B.Sc. thesis: NEURAL NETWORKS : A MATHEMATICAL OVERVIEW (supervisor: Dr. J.R. Wedagedara)
• project abstract(PDF)
The thesis discusses some of the important neural network models and their mathematical aspects: namely, feedforward models-perceptron, multilayer networks and associative memories-linear feedforward memories, bidirectional associative memories and hopfield networks. It explores the applicability of a special case of heteroassociative model - constructed using Multiple training encoding strategy and operating in feedforward mode to one of the real-world problems - pattern recognition, Sinhala handwritten character recognition, in particular.
Keywords: Neural Networks, Perceptron, Backpropagation, Hopfield networks, Bidirectional associative memories, Hand-written character recognition.

• project summary(PDF)

Further works

• It is interesting to view the superposition of the target character among the trained samples: For an example the Sinhala letter "a" was stored among other trained characters like this
  
         

• What the hidden layer neurons learn:
  
  The hidden neurons of a multi-layer network with a single hidden layer can only implement linear decision regions. This fact was observed during a simulation suggested by Hassoun.
  With the sample points of the two classes taken as this :
  
  
  and trained using a logistic function an 8 hidden layer nodes at their best were only able to realize linear decision boundaries.
  
  
  
  


• What the hidden layer nodes of a two hidden layer network learns:
  Hidden nodes in a 8-4-1 network learns arbitrary decision boundaries.
  
  

Back to top

• RSA crypto system (HTML)
• Rabin Test (HTML)
• Solving linear congruences (HTML)
• Egyptian exponentiation (HTML)
• Euler Phi function - Visualization (PDF)

Back to top