Handwritten digit recognition using trace neural network with EKF training algorithm
K. Wang, Chang Sheng-Jiang, Chi-Sing Leung · 2003
The authors propose combining the trace learning rule with an on-line dual extended Kalman filter algorithm for invariance extraction and recognition of handwritten digits. In order to reduce the sensitivity of the extracted invariance to samples with large variance, a novel activation function is proposed to replace the traditional sigmoid activation function. Computer simulations show that both the learning speed and the recognition rate are improved.