The learning of multi-output binary neural networks for handwritten digit recognition

A new learning method of multi-output binary neural networks (BNN) is proposed for handwritten digit recognition based on our simulated light sensitive model. The new teaming algorithm guarantees convergence for any binary-to-binary mapping including these multi-output cases, and learns much faster than the backpropagation learning algorithm. Neurons in the BNN employ a hard-limiter activation function and integer weights, thus greatly facilitating hardware implementation of BNN using current digital VLSI technology.