N-TORC

we describe N-TORC, a tool flow for generating a candidate set of neural network models for a target dataset that achieve the highest possible accuracy for a given resource cost while meeting a real-time latency constraint. We evaluate this approach using a benchmark structural state estimation dataset, DROPBEAR, but in principle, the approach can be used for any dataset that can be trained with a model whose parameters can fit in the BRAM of an embedded-class FPGA.