Fig. 4: Discretization enables knowledge-distillation on Mackey-Glass.
From: A predictive approach to enhance time-series forecasting
a After generating the chaotic trajectory, every scalar target Xt+ℓ is quantized into one of C equal-width bins. The regression problem is thus recast as C-way classification, so the teacher and student can exchange soft logits of identical dimensionality-an essential requirement for the KL-distillation term in FGL. b Teacher performs next-step prediction. c Baseline performs a 5-step forecast without future guidance. d FGL-trained student forecasting the same horizon. By aligning its logit distribution with the teacher’s near-future logits, the FGL student captures neighborhood information in bin space, yielding a visibly lower MSE and a smoother reconstruction.
