On: Implementing Neural Networks on a "10-cent" RISC-V...

[cpldcpu]

The latter one. The network is trained in full precision (this is required for the gradient calculation), but the weights are nudged towards the quantized values.