Simultaneously Optimizing Weight and Quantizer of Ternary Neural Network using Truncated Gaussian Approximation