Google launched Neural Structured Learning, an open source framework that uses Neural Graph Learning method to train neural networks with graphs and structured data.
Most of the file AI models process are unstructured data like videos and images. But other kinds of files can be used for machine learning projects. Tables of structured information are useful in training newly-built AI models in efficient pattern recognition.
“Leveraging structured signals during training allows developers to achieve higher model accuracy, particularly when the amount of labeled data is relatively small,” Google engineers Da-Cheng Juan and Sujith Ravi wrote in a blog post for Neural Structured Learning.
Neural Structured Learning let’s programers (and non programers) incorporate structured data into a project, in some cases with as little as five lines of code. All that is needed as a prepared AI model, and to provide the training records and structure which the records are formatted.
The AI models scientists use for genomics and molecular research often require structured data inputs. Neural Structured Learning can make models for computer vision, perform NLP, and run predictions from graphical datasets like medical records or knowledge graphs.
The framework allows the creation of adversarial examples. Adversarial examples are files like a photo with manipulated pixels, that a human processes easily, but can corrupt an AI’s processing results. These kind of records used in machine learning model development train software to fend off potential attacks.
“Empirically, models trained without adversarial examples suffer from significant accuracy loss (e.g., 30% lower) when malicious yet not human-detectable perturbations are added to inputs,” says the blog post.
Neural Structured Learning is likely to prove particularly useful for teams working on AI with sensitive or publicly-facing inputs that must address the possibility of malicious input data.
Header Image: “NYC FLOW” by Danil Krivoruchko