#discriminative-fine-tuning

ULMFiT fine-tunes each layer with its own learning rate (discriminative fine-tuning). It first picks the rate $\eta^L$ of the last layer and then sets $\eta^{l-1} = \eta^{l}/2.6$ for each lower layer. If $\eta^L = 0.01$, what learning rate does layer $L-3$ get?