To understand AutoML, which is short for automated machine learning, let’s briefly look at how it was built. If you've worked with ML models before, you know that training and deploying ML models can be extremely time consuming, because you need to repeatedly add new data and features, try different models, and tune parameters to achieve the best result. To solve this problem, when AutoML was first announced in January of 2018, the goal was to automate machine learning pipelines to save data scientists from manual work, such as tuning hyperparameters and comparing against multiple models. But how could this be done? Well, machine learning is similar to human learning. It all starts with gathering the right information. For AutoML, two technologies are vital. The first is known as transfer learning. With transfer learning, you build a knowledge base in the field. You can think of this like gathering lots of books to create a library. Transfer learning is a powerful technique that lets people with smaller datasets, or less computational power, achieve state-of-the-art results by taking advantage of pre-trained models that have been trained on similar, larger data sets. Because the model learns via transfer learning, it doesn’t have to learn from scratch, so it can generally reach higher accuracy with much less data and computation time than models that don’t use transfer learning. The second technology is neural architecture search. The goal of neural architecture search is to find the optimal model for the relevant project. Think of this like finding the best book in the library to help you learn what you need to. ​​AutoML is powered by the latest machine-learning research, so although a model performs training, the AutoML platform actually trains and evaluates multiple models and compares them to each other. This neural architecture search produces an ensemble of ML models and chooses the best one. Leveraging these technologies has produced a tool that can significantly benefit data scientists. One of the biggest benefits is that it’s a no-code solution. That means it can train high-quality custom machine learning models with minimal effort and requires little machine learning expertise. This allows data scientists to focus their time on tasks like defining business problems or evaluating and improving model results. Others might find AutoML useful as a tool to quickly prototype models and explore new datasets before investing in development. This might mean using it to identify the best features in a dataset, for example. So, how does AutoML work exactly? AutoML supports four types of data: image, tabular, text, and video. For each data type, AutoML solves different types of problems, called objectives. To get started, upload your data into AutoML. It can come from Cloud Storage, BigQuery, or even your local machine. From there, inform AutoML of the problems you want to solve. Some problems may sound similar to those mentioned in pre-built APIs. However the major difference is that pre-built APIs use pre-built machine learning models, while AutoML uses custom-built models. In AutoML, you use your own data to train the machine learning model and then apply the trained model to predict your goal. For image data: You can use a classification model to analyze image data and return a list of content categories that apply to the image. For example, you could train a model that classifies images as containing a dog or not containing a dog, or you could train a model to classify images of dogs by breed. You can also use an object detection model to analyze your image data and return annotations that consist of a label and bounding box location for each object found in an image. For example, you could train a model to find the location of the dogs in image data. For tabular data: You can use a regression model to analyze tabular data and return a numeric value. For example, you could train a model to estimate a house’s value or rental price based on a set of factors such as location, size of the house, and number of bedrooms. You can use a classification model to analyze tabular data and return a list of categories. For example, you could train a model to classify different types of land into high, median, and low potentials for commercial real estate. And a forecasting model can use multiple rows of time-dependent tabular data from the past to predict a series of numeric values in the future. For example, you could use the historical plus the economic data to predict what the housing market will look like in the next five years. For text data: You can use a classification model to analyze text data and return a list of categories that apply to the text found in the data. For example, you can classify customer questions and comments to different categories and then redirect them to corresponding departments. An entity extraction model can be used to inspect text data for known entities referenced in the data and label those entities in the text. For example, you can label a social media post in terms of predefined entities such as time, location, and topic, etc. This can help with online search, similar to the concept of a hashtag, but created by machine. And a sentiment analysis model can be used to inspect text data and identify the prevailing emotional opinion within it, especially to determine a writer's comment as positive, negative, or neutral. And finally, for video data: You can use a classification model to analyze video data and return a list of categorized shots and segments. For example, you could train a model that analyzes video data to identify whether the video is of a soccer, baseball, basketball, or football game. You can use an object tracking model to analyze video data and return a list of shots and segments where these objects were detected. For example, you could train a model that analyzes video data from soccer games to identify and track the ball. And an action recognition model can be used to analyze video data and return a list of categorized actions with the moments the actions happened. For example, you could train a model that analyzes video data to identify the action moments involving a soccer goal, a golf swing, a touchdown, or a high five. In reality, you may not be restricted to just one data type and one objective but rather need to combine multiple data types and different objectives to solve a business problem. AutoML is a powerful tool that can help across these different data types and objectives.
