Title:

Polarity Graphs: A Structure for Pure Dataflow Computation

Abstract:

Graphs are a natural way to represent computation. Many such representations are intensional: they unfold through conditional branching. In some domains, such as digital circuits, machine learning inference, and scientific workflows, an extensional representation of the dataflow is sufficient. What kind of graph can provide a domain-neutral and mathematically clean model of pure dataflow?

In this talk I will introduce polarity graphs, my answer to that question. A polarity graph is, informally, a typed bipartite graph with two kinds of nodes: values and operations. The graph is equipped with an interface that its boundary must realize exactly. I will explain how this structure captures pure dataflow in a way that is adaptable across domains.

Speaker: Frej Dahlin