Slide 1Project GoalMotivationMethodologyNeural Network?Slide 6Slide 7DiscussionDynamic Hand Written Character RecognitionShi-Ting ZhouProject Goal•Devise a new approach in recognizing hand written characters without training process in respective classical neural network approaches.•Validate the proposed technique by testing on hand written digitsMotivation•Classical ANN does not possess the merits of biological neural systems•Real biological neural systems are dynamical•Classical ANN approaches ignore 2-D information of inputsMethodology•Devise a simulator capable of simulating the dynamics of elastic body in attracting force field•Build templates of written digits•Test on the MNIST DATABASE of handwritten digits (“http://yann.lecun.com/exdb/mnist/”)Neural Network?•Can be implemented in the form of recurrent neural network.Discussion•Potential usage on other type of recognition task•Can have a 3-D version. 1-D lines become 2-D manifolds•Problem: Computationally intensive if have a lot of templates•Solution: Targeting at smaller and repeating features instead of entire
View Full Document