AutoML is an important technology that simplifies the building, training, and deploying of machine learning models. In short, it strips away the complexity in the end-to-end workflow so even novice users can experience the power of machine learning. Startups, large enterprises, and communities around the world are developing feature-rich tools that are easy to use, in some cases only requiring a few lines of code to operate, like AutoGluon, one of the most popular tools in this category.
The promise of AutoML makes it sound like fiction, but it’s not. One day in the future, it will be everywhere, not only in the cloud but also on cars, phones, websites, applications, and much more, each piece playing its part. And it will require minimal human intervention to build, train, and deploy models on the fly. Today, some tools like AutoGluon are able to prepare the data, cleanse it, perform feature engineering, algorithm selection, hyperparameter tuning, model selection, training, and much more, courtesy of Amazon, the organization behind the tool.
AutoGluon can “train state-of-the-art machine learning models for image classification, object detection, text classification, and tabular data prediction” without the need of being an expert. There are other popular open-source AutoML tools like Auto-sklearn, NNI, and Google Model Search. Here’s a list of tools that have been identified in the marketplace. Some of them continue to evolve in maturity and have large communities behind them.
AutoML Tools
- AutoGluon
- TransmogrifAI
- Auto-sklearn
- Neural Network Intelligence (NNI)
- Auto-Keras
- TPOT
- AutoWeka
- Model Search
Some of the most difficult tasks for the non-data scientist in working with machine learning are data preparation, algorithm selection, model selection, and hyperparameter tuning. For the latter, it’s just endless possibilities. AutoML figures all that out in minutes.
Open-Source AutoML List
| Name | Type | Notes | Features |
|---|---|---|---|
| AutoGluon | Works with deep learning and classical ML | Supports text, object, and tabular data | Auto hyperparameter tuning, model selection, architecture search |
| TransmogrifAI | Accelerates ML developer productivity | Works with structured data | Algorithms help with feature engineering, feature and model selection |
| Auto-sklearn | Out of the box supervised machine learning | Comes with 15 classification and 14 feature preprocessing algorithms | Searches for right dataset and optimizes hyperparameters |
| NNI | Manages AutoML experiments | Searches for the best neural architecture and hyperparameters | Automates feature engineering, hyperparameter tuning, model compression, and NAS (search) |
| Auto-Keras | Works with deep learning, classification, and regression | Tightly integrated with TensorFlow | Finds the best hyperparameters and model architecture |
| TPOT | Initially developed for the science community | Optimizes ML pipelines | Explores thousands of possible pipelines to identify the best one |
| Auto-Weka | For data prep, classification, regression, clustering, and visualization | Helps identify best algorithms and hyperparameter for given application | Collection of algorithms for data mining |
| Google Model Search | Helps with classification problems | Runs out of the box, searching and comparing models | Model architecture search |
While some AutoML tools provide a narrower set of out-of-the-box functionality, others like AutoGluon are comprehensive. In three years, our guess is AutoML will go mainstream.