Azure AutoML is a cloud-based service that can be used to automate building machine learning pipelines for classification, regression and forecasting tasks. Its goal is not only to tune hyper-parameters of a given model, but also to identify which model to use and how to pre-process the input dataset.

AutoML uses Bayesian optimization to identify better hyper-parameters than human experts. It also uses collaborative filtering (Probabilistic Matrix Factorization) to search for the most promising data transformation pipelines efficiently, based on a database that is built by performing millions of different pipeline evaluation experiments on many datasets. This database helps in finding the good solutions for new datasets quickly.

Automated Machine Learning is certainly a tool that greatly simplifies the operations required for training ML models.

Thanks to Automated ML the inexperienced users can obtain a model having satisfactory performance with minimal effort, unknowingly following the best practices of the most experienced data scientists.

Even the most experienced data scientists can benefit from using Automated ML. They have the possibility to focus their attention and exploit their expertise on the preparation of the training dataset (choosing the best dataset shape that can improve the ML results), leaving the repetitiveness of the operations necessary to train the model to the Automated ML.

There are more than one algorithm supported by Azure AutoML depending on the type of problem to be addressed.

Below is an exhaustive list of supported algorithms at the time of this writing. If a model has to be run on a variety of platforms and devices, converting it in the ONNX format could be the right solution. In that case, only those algorithms indicated with an * are able to be converted to the ONNX format.

Very often, analysts working with classical machine learning do not have sufficient expertise to make the best use of neural networks. A tool that would allow to do automatic training of neural networks applied to classical problems such as classifications and regressions would be a godsend for many users.

AutoML doesn’t yet provide Deep Neural Network algorithms among those supported for training ML models (except ForecastTCN for time-series forecasting, as mentioned before), even if Microsoft is already working on an open source project called Neural Network Intelligence that aims to automate Feature Engineering, Neural Architecture Search, Hyper-parameter Tuning and Model Compression.

Azure AutoML is a relatively young tool, on which the development team is making continuous improvements day by day. So it is expected that some missing features will be implemented in the next milestones.

QUESTION I : Could we call Azure AutoML user friendly?
QUESTION II : Does Azure AutoML need what if analysis?

REFERENCE : A Review of Azure Automated Machine Learning