Fig. 6
From: Joint high-dimensional soft bit estimation and quantization using deep learning
Comparing the two-stage training of EQ-Net and a naïve supervised baseline that jointly trains the estimation encoder and the decoder from scratch, in a single stage. In both cases, the networks have the same architecture and initial weights, and the same number of gradient updates is performed. The plot includes shaded standard deviation areas over ten runs. (Left) Validation error (soft bit distortion). (Right) Validation coded block error rate with models taken after 500 epochs