Extend recursive-relation detection to complex machine-learning algorithms

Extend the ProvQML recursive-relation detection algorithm to support more complex machine-learning algorithms, such as neural networks, by accounting for additional base cases per recursive relation and deeper recursive dependencies.

Background

ProvQML identifies mutually recursive and non-linearly recursive relations in machine-learning queries using a single-pass scan of each recursive case. The paper notes that this procedure is sufficient for the fixed recursive structure of the machine-learning queries considered, which contain a primary recursive relation and auxiliary relations for prediction and gradient computation.

The authors explicitly leave open the extension of this detection procedure to more complex machine-learning algorithms, including neural networks. Such an extension would require the algorithm to recognize additional base cases for each recursive relation and to inspect deeper recursive dependencies beyond the fixed structure supported in the paper.