Flattened one-bit stochastic gradient descent: compressed distributed optimization with controlled variance | Synapse