通过精度分配方法,实现神经网络中所有参数的最小化,从而实现固定点训练。针对 CIFAR-10,CIFAR-100 和 SVHN 数据集,对四个网络进行实验验证,证实此方法具有接近最优的精度分配,可以与其他固定点神经网络设计相比较。(The precision assignment methodology reduces the complexity of fixed-point training for neural networks, and its optimality is validated empirically for various datasets and network designs)