Data-flow models of computation define programs as a graph of computation nodes connected by communication edges. Such models have many advantages among which a formal deterministic semantics, static analyses, and highly parallel implementations. Up to now, their main application domains have been multimedia applications and signal processing. However, they also seem particularly well suited for the efficient implementation of neural networks. The goal of the DF4DL exploratory action is to study the potential of data-flow models for implementing neural networks, in particular when such networks require dynamicity (changes and reconfigurations). We expect advances in the form of better computation performances and a lower energy consumption of deep learning applications.
Inria teams involved