Two trends rose in acceptance in 2020:
- Knowledge Transfer, i.e., nobody trains neural networks from scratch, and
- AutoML, i.e., all traditional models tried with hyperparameter tuning at once, and the most successful model automatically chosen for further use.
This article is about the second one.
The Year 2000
MS SQL Server 2000 appeared. One of the added components: SQL Analysis Services. One of its parts: Data Mining tools. An ancient term for Machine Learning. In general, many ML algorithms and fully-connected Neural Networks are operating over data found in the DB. Cool things:
- You didn’t have to export the data to train a model with them
- Traces of semi-automatic feature engineering
- Models are retrained in a scheduled manner
- Models lived within the DB
- Models are asked for inference via SQL-like queries, with commands added to the TSQL.
One important note: no progress here since the very beginning. They are today as they were 21 years ago. Not that good sign for them, even they looked promissing they.
The year 2020
Mindsb: The next step. Cool things:
- Everything just said about MS SQL Server 2000, in the text above,
- Any major DB supported,
- Automated feature selection and engineering
- The model selected with hyperparameters fine-tuned via the AutoML approach.
I had a video meeting with the guys who made it. They executed a few real-live inferences, which looked really cool. If I need something similar, I will use this for sure.
What happened similarly in-between: I haven’t found any traces.
Midnsdb is opensource and free.