Applying linear subspace constraints in the shuffling forward model enables reconstruction of signal dynamics through reduced degrees of freedom. Other techniques train auto-encoders to learn latent representations of signal evolution to apply as regularization. This work inserts the decoder portion of an auto-encoder directly into the shuffling forward model to reduce degrees of freedom in comparison to linear techniques. We show that auto-encoders represent fast-spin-echo signal evolution with 1 latent variable, in comparison to 3-4 linear coefficients. Then, the reduced degrees of freedom enabled by the decoder improves reconstruction results in comparison to linear constraints in simulation and in-vivo experiments.