Why Proper Training of Deep Learning Inspection Applications is Crucial for Success

As with any new factory automation technology, there are considerations and trade-offs that come with its adoption and rollout. While deep learning-based machine vision promises to solve many complex factory applications, it’s not by any means a panacea or silver bullet. That’s why setting proper expectations for what it can do is important to any project; knowing what it takes to create an application upfront is key.

One of the biggest considerations when it comes to building deep learning-based inspection applications is training the application. That’s because deep learning-based applications are not explicitly programmed; rather, they are trained from reference images to spot anomalies that fall outside the accepted range.

A good analogy for why proper training of an artificial intelligence application is important, argues Nate Soares, a former Google software engineer who runs the Machine Intelligence Research Institute, comes from the Disney animated feature Fantasia – specifically the Sorcerer’s Apprentice scene.

“The problem Mickey faces when he enchants a broom to help him fill a cauldron isn’t that the broom rebels or acquires a will of its own, but that it fulfills the task it was given all too well,” Soares said in an interview in 2018. “He wants the cauldron full, and overflowing the workshop is a great way to be extra confident that the cauldron is full (and stays full). Mickey successfully ‘aimed’ his AI system, but things still went poorly for Mickey.”

Why training a deep learning-based application matters

In other words, the result or output of a deep learning-based application, without proper training of that system, might be unexpected and that’s not good when a manufacturer needs reliable inspection results. In the case of factory automation, application engineers need to understand that a well-trained deep learning-based application requires a comprehensive set of training images that represent a range of defects and/or acceptable part variations to perform well in production.

Those images also need to be acquired under manufacturing lighting and part presentation conditions. This is essential for any deep learning project to become successful.

Qualifying a deep learning vision solution is an iterative process that requires the system to be installed on a production line. And, unlike traditional machine vision systems, training and validation of the images for deep learning must be done during the development phase – it can’t wait until factory acceptance testing. Deep learning requires a great number of samples to train with, which could require time to capture the representative set of images needed to train a well-performing deep learning tool.

“Sometimes, deep learning systems perform well in the lab but struggle when deployed in the production environment,” says Grace Lee, Senior Product Marketing Manager for AI at Cognex. “User frustration stems from the underappreciated differences between deep learning solutions and the more familiar traditional machine vision.”

Lee works on Cognex’s artificial intelligence team that helps deliver deep learning factory automation solutions. They’ve recently implemented those deep learning algorithms into the software that powers the machine vision cameras in order to solve more complex and challenging inspection problems.

These tools help her customers make more accurate and scalable decisions when inspecting for defects or anomalies. She likens the work Cognex does with artificial intelligence, and specifically deep learning, as similar to how other AI-based tools help people make smarter decisions when purchasing flights, buying stocks, or recommending new music to listen to. Instead, Cognex’s deep learning solutions help manufacturers make smarter decisions about quality inspections that have otherwise been performed by humans because they were too difficult to automate.

“The application of AI in a factory setting isn’t a far-off promise,” says Lee. “It’s very much solving actual challenges now. But engineers need to think through the ways in which they set up projects, train them to be effective, and ultimately execute on them.”

To learn more, download the Getting Started with a Deep Learning Factory Automation Project eBook.