This article gives a quick overview of Visual AI in DataRobot: what it is and how to use it to rapidly build and deploy models that take into account information extracted from image data. This video provides an end-to-end walkthrough for Visual AI.
Visual AI provides the ability to include images in your Supervised Machine Learning pipeline (Figure 1). Similar to other DataRobot projects, Visual AI projects deliver both deployable models and associated model insights. One such insight shows which aspects of the input images the model is focusing on when attempting to discriminate between classes.
Figure 1. A sample dataset with an image variable type
Visual AI has a wide variety of highly efficient, state-of-the-art, deep learning featurizers including Squeezenet, resnet50, xception, efficient-net, and some others. Each of these featurizers convert an input image to a vector of numbers (Figure 2). You can apply Visual AI to any kind of problem because it combines features from various layers of a given featurizer and is not limited to using the output of that pre-trained featurizer.
Figure 2. A simplified version of the deep learning architecture used in the Visual AI platform
Although this may appear very complex, when using Visual AI in your projects you get to focus on your business problem rather than wrestling with deep learning concepts and code intricacies, thinking about how to provision expensive GPUs, or sweating over how to manually label millions of images before they can be used in the model building pipeline. The DataRobot deep learning architecture requires only a few hundred images to build a functional model.
Creating projects with images follows the same end-to-end workflow as used for DataRobot projects without images (Figure 3).
You submit your data by dragging and dropping it into the platform or using the AI Catalog. The data can have a mix of data types as long as one is of image type.
Then you configure the settings in DataRobot based on your project needs.
Press Start to initiate the model building process.
Make predictions after thoroughly scrutinizing your preferred model using available model insights tools within the DataRobot platform.
Figure 3. The end-to-end workflow of a Visual AI project
Visual AI has extra tools that are specific to the image data type which were created to enhance model insights.
Image Activation Maps allow you to see sample locations in the image that the model is using to make decisions. (Notice the color highlights showing areas of high and low activation (Figure 4).)
Figure 4. Image Activation Maps showing areas of high and low activation
Image Embeddings allow you to visualize a sample of images projected from their original N-dimensional feature space to a new 2-dimensional feature space. This makes it easy to see what images are considered similar (Figure 5).
Figure 5. Image Embeddings showing what images were clustered together due to similar image features