TrainDB’s Main R&D Topics

  1. TrainDB: an ML-model based approximate query processing engine

    • SQL-like approximate query language

    • Approximate query processing using synopsis data that are synthesized by ML models

    • Approximate query processing using ML inference models

    • Various DBMS data sources support via extensible data source adapters

  2. ML model library for approximate query processing

    • Synopsis generative ML models + inferential ML models

      • Synopsis generative ML models: GAN-based models(e.g., TableGAN, OCT-GAN), score-based generative models

      • Inferential ML models: mixture density networks(MDN), relational sum-product networks(RSPN)

    • Error estimation for approximate query evaluation

    • Continual learning to update base table changes

  3. Cloud ML model serving framework

    • A framework for training/serving ML models in remote GPU servers

    • ML model registry/training/serving support

  4. Visual Exploratory Data Analysis Support Tools for TrainDB

    • Approximation query result visualization for exploratory data analysis

    • Visual OLAP analysis support for multi-dimensional data analysis