8 Deep Learning in Non-Compositional Sketches

Deep Learning in Non-Compositional Sketches, or DLINCS, turns the structural language of the preceding chapters into a computational discipline for layered models. A deep model is not treated as one opaque parametrized function. It is presented as a typed learning sketch whose serial and parallel compositions are explicit, whose equations state the intended architecture, and whose tangent and reverse computations retain that structure.

¹1. We use Deep LINCS for the computational realization and \(\mathrm{dlincs}\) in formulas. Its role is to compile declared obstructions into trainable, auditable modules. ↩