Universally Quantized Neural Compression

            Eirikur Agustsson              Lucas Theis
             Google Research              Google Research
           eirikur@google.com             theis@google.com

                         Abstract
       A popular approach to learning encoders for lossy compression is to use additive
       uniform noise during training as a differentiable approximation to test-time quan-
       tization. We demonstrate that a uniform noise channel can also b ...