Automated Machine Learning has come a long way since Google Brain introduced NAS in early 2017. Amazon’s AutoGluon, Google’s AutoMLZero, Salesforce’s TransmogrifAI, the AutoML features of Azure, H2O, Scikit-learn, Keras & others (TPOT, DataRobot, etc.) are witnessing increased adoption.

Modern AutoML systems generally focus on hyper-parameter optimization (HPO), neural architecture search (NAS), model selection & compression, and to a certain extent, meta-learning. They also address tasks like basic feature engineering, simple visualizations, data augmentation & data imputation.

However, most AutoML tools suffer from several limitations which impede large-scale adoption today.
• AutoML for Computer Vision, Deep Learning & complex NLP are in early stages of evolution.
• Integration with large DevOps structures and debugging/error handling are practically difficult.
• Performance issues & system failures for compute-intensive workloads like HPO & NAS are common.
• Issues like configurability, explainability, fairness & training time estimation are inadequately addressed.
These shortcomings will get addressed as time progresses. The future is in Low-Code, No-Code & Visual AutoML, which will truly help in democratizing Artificial Intelligence.