In this paper, the drawback is pointed out that the learning algorithm em of random neural network sometimes converges to local minimum. A new hybrid learning algorithm HRem, which combines algorithm em and the random optimization algorithm presented by Dr. Solis and Wets, is presented for 3-layer random perception. It is theoretically proved that algorithm HRem can globally converge to the minimum of Kullback-Leibler difference measure. This theoretical result has important significances for further research on algorithm em.