Deep Learning Provenance Data Integration: a Practical Approach

Abstract

A Deep Learning (DL) life cycle involves several data transformations, such as performing data pre-processing, defining datasets to train and test a deep neural network (DNN), and training and evaluating the DL model. Choosing a final model requires DL model selection, which involves analyzing data from several training configurations (e.g. hyperparameters and DNN architectures). Tracing training data back to pre-processing operations can provide insights into the model selection step. Provenance is a natural solution to represent data derivation of the whole DL life cycle. However, there are challenges in providing an integration of the provenance of these different steps. There are a few approaches to capturing and integrating provenance data from the DL life cycle, but they require that the same provenance capture solution is used along all the steps, which can limit interoperability and flexibility when choosing the DL environment. Therefore, in this work, we present a prototype for provenance data integration using different capture solutions. We show use cases where the integrated provenance from pre-processing and training steps can show how data pre-processing decisions influenced the model selection. Experiments were performed using real-world datasets to train a DNN and provided evidence of the integration between the considered steps, answering queries such as how the data used to train a model that achieved a specific result was processed.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 9 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.