Google LLC today open-sourced Neural Structured Learning, a framework for its popular TensorFlow artificial intelligence development toolkit that will enable developers to train models with structured and “adversarial” data.

Most of the files that AI models are used to process, such videos and images, are technically unstructured data. But other types of files still have their place in machine learning projects. Structured information organized as a graph is helpful for training newly-built AI models how to recognize patterns efficiently. 

“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 explained in the blog post today for Neural Structured Learning.

Beyond being a useful training method, processing structured data is an integral task for certain types of machine learning software. The AI models that scientists employ in genomics and molecular research often take structured graph data as input. The same goes for certain types of natural-language processing algorithms.

Neural Structured Learning allows developers to incorporate structured data into a project with just a few lines of code. They only need to get their AI model ready, provide the training records and specify the structure according to which the records should be organized.

The framework also lets developers specify “implicit” structures to create so-called adversarial examples. Adversarial examples are malicious files, such a photo with manipulated pixels, that don’t seem different to a human but can trip up an AI and corrupt processing results. Throwing such records at a machine learning model during development teaches the software how to fend off 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,” Google’s Da-Cheng Juan and Sujith Ravi wrote. As a result, Neural Structured Learning may prove particularly handy for AI teams working on sensitive or publicly-facing services that must address the possibility of manipulation attempts.

Image: Google