Chris Fregly - Data Science on AWS: Implementing End-to-End, Continuous AI and Machine Learning Pipelines

Automated machine learning (AutoML) commonly refers to the effort of automating the typical steps of a machine learning pipeline shown in .

Figure 1-1. Typical machine learning pipeline.

Machine learning practitioners spend a lot of time building and managing such pipelines. They need to prepare the data and decide on the framework and algorithm to use. Seasoned data scientists use years of experience and intuition to choose the best algorithm for a given dataset. In an iterative process, ML practitioners try to find the best performing model configuration called hyper-parameters. Unfortunately, there is no cheat sheet either for choosing any of these parameters. We still need experience, intuition, and patience to run many experiments and find the best hyper-parameters for our algorithm and dataset.

What if we could just use a service that automatically finds and trains the best model for our dataset and deploys the model to production with a single click? Amazon SageMaker Autopilot offers us exactly this functionality. Autopilot simplifies the model training and tuning processing by handling many aspects of the model development life cycle (MDLC) including feature transformation, algorithm selection, model training, tuning, and deployment.

Simply point Autopilot to your dataset - and out comes a set of fully-trained and optimized predictive models. Autopilot explores many algorithms and configurations based on many years of AI and machine learning experience at Amazon. The model candidates are summarized by Autopilot through a set of generated Jupyter notebooks and Python scripts. You have full control over these generated notebooks and scripts. You can modify them, automate them, and share them with colleagues. You can select the top model candidate based on your desired balance of model accuracy, model size, and prediction latency. Lets dive deeper into the process of automated machine learning with Autopilot.

Automated Machine Learning with Autopilot

Autopilot is the name of Amazon SageMakers AutoML service. You simply provide your raw data in a S3 bucket, for example in the form of a tabular CSV file, and tell Autopilot which column you want to predict. As the name implies, Autopilot then does the rest automatically.

Note

S3 is Amazons Simple Storage Service. S3 provides a simple web service interface that you can use to store and retrieve any amount of data. We will discuss this service in more detail in the next chapter.

Autopilot uses automated machine learning to analyze the data and identifies the best algorithm and model configuration for your data as shown in .

Figure 1-2. Amazon SageMaker Autopilot.

You can tell Autopilot how many model candidates to explore. In the process of building those candidates, Autopilot tries different algorithms and algorithm settings. Autopilot also applies all needed data transformations to your data to optimize the input for each algorithm. The algorithm, configuration, and data transformation code are then combined into a single ML pipeline definition. The most promising pipelines are selected by Autopilot and used to find the best performing model. Lastly, Autopilot shares the results in a model leaderboard. You can use the best performing model as a baseline and optimize the model even further. A second option is to simply deploy the model and start predicting.

Another highlight of Autopilot is the fact that it provides full visibility into each of those steps and shares all code needed to reproduce the results. AWS calls this a white-box approach. This white-box approach to AutoML is very unique. Lets explore the white-box vs. black-box approach to AutoML a bit further.

Understand Autopilots White-Box Approach to AutoML

In a black-box approach as shown in , you dont have control or visibility into the chosen algorithms, applied data transformations, or hyper-parameter choices. You point the AutoML service to your data and receive a trained model.