Sampled Softmax with Random Fourier Features

  Ankit Singh Rawat, Jiecao Chen, Felix Yu, Ananda Theertha Suresh, and Sanjiv Kumar
                Google Research, New York
    {ankitsrawat, chenjiecao, felixyu, theertha, sanjivk}@google.com

                            Abstract
       The computational cost of training with softmax cross entropy loss grows linearly
       with the number of classes. For the settings where a large number of classes
       are involved, a common method to speed u ...