Pipelined Training with Stale Weights in Deep Convolutional Neural Networks

Figure 6

Inference accuracy vs % of stale weights.