What is the Difference Between AI, Machine Learning, and Deep Learning for Industrial Automation Inspections?
It can often be hard to separate fact from marketing buzz when it comes to artificial intelligence (AI), especially in factory automation inspections. In many ways, AI is already a part of our everyday lives and in many other respects the technology is still a futuristic concept.
The spam filters on your email client leverages AI so you don’t have to read those unwanted junk emails. On the other hand, a robot that could pass the Turing Test such as C-3PO from Star Wars exists only in sci-fi movies and TV shows. Over the past few years, the practicality of using AI has exploded thanks to the rapid cost improvements for compute power, and cloud storage, as well as the increase in production of data from images, texts, software transactions, and more.
Terms like AI, machine learning, and deep learning are often used interchangeably with little to no explanation or context as to what each term means. By understanding the different flavors of artificial intelligence, including machine learning and deep learning, it’ll be easier to understand how this technology is and isn’t helping manufacturers and factories today.
Artificial Intelligence – Programming Tasks with Logic
To start, artificial intelligence is a subset of computer science wherein computer systems perform human-like tasks (e.g. image classification, speech recognition, language translation) as good as or better than humans. The logic concepts were born in the pioneering computer science days of the 1950s.
AI is loosely segmented in two ways: narrow AI and general AI. “General AI” is that futuristic concept of robots acting and thinking like humans. Maybe one day, we’ll have to deal with sentient robots, but, for the purposes of this article, we’ll be focusing only on “Narrow AI”, which is any computer system designed to perform specific human-like tasks.
AI is really the discipline of creating smart algorithms. In the early days, AI was just a set of programmed computer instructions. Now, it can be anything from a complicated set of logic to a self-learning algorithm that generates outcomes based on reference examples and with minimal human programming.
The operation of traffic lights is an example of AI in the real world. What once required humans to change from red to green can now be done with smart logic and programming to allow one light to stay green for 45 seconds and then turn red. This is known as fixed time control. There are other approaches to programming a network of traffic lights, such as coordinated control which attempts to provide drivers with a long progression of green lights. Fundamentally, the operation of streetlights is just programming them to achieve a specific task, so humans no longer have to manually do it.
Machine Learning – The Application of Artificial Intelligence
Machine learning was developed as a subset of AI and is considered a technique for achieving AI. Machine learning is the practice of using algorithms to empowering computer systems with the ability to learn from data and make decisions.
The decision-making algorithms over the years have included decision trees, cluster analysis, reinforcement learning, and Bayesian networks, amongst many others. Referring back to our traffic light example, machine learning algorithms could be used to help determine a pattern for the optimal length of time for switching from red to green based on time of day or traffic congestion. This is something cities such as Las Vegas, NV are already piloting with the hopes that machine learning will help play a role in reducing traffic congestion by 40%. Machine learning helps cities go one step beyond just programming their traffic lights, but to leverage generated data from cars to make optimal decisions on that programmed logic.
Finally, machine vision is one of the best uses of machine learning. By leveraging image data captured by cameras and a diverse set of algorithms applied to those images such as classifiers, location tools, and even optical character recognition it’s possible for machine vision software to determine the presence or absence or a part, for example, gauge the width between two edges, or identify a string of characters on a tire.
Here at Cognex, we refer to machine vision as a traditional or rule-based approach to solving inspection challenges. Our rules, which are technically algorithms, are really just software tools used by humans to program specific tasks such as finding two edges of a part and then determining the width between those edges. We don’t think of them as machine learning or AI even if they might fall under that umbrella.
Deep Learning – The Next Evolution of Inspection
Deep learning algorithms are the latest subset of artificial intelligence to gain prominence thanks to continued advances in technology. Deep learning builds off of the advances made under machine learning but with a few key differences.
Instead of relying on humans to program tasks through computer algorithms, deep learning reaches outcomes through an example-based approach that mimics human learning. By leveraging neural networks, deep learning-based inspection applications make connections and spots patterns from massive data sets.
For example, let’s say a manufacturer wants to detect defects in the products they make. One way to approach this would be via traditional machine vision. With traditional vision, engineers would have to explicitly program the inspection to account for the millions of variations that could happen: the size and type of defect, the location of the defects, and so on. That becomes a very time-consuming application to both maintain and program because of the inherent variations.
With a deep learning-based approach, the algorithm takes the examples provided by the user and automatically creates an understanding of the part being inspected. By creating an inspection that learns what a good part looks like, even accounting for slight variations, the solution can then flag when something looks amiss such as a scratches, foreign objects, or other visual defects.. Users can then improve the solution by providing more data for the tool to learn from. The more data the deep learning application has, the better it will perform over time to spot anomalies.
The Future of Factory Automation Inspections
While artificial intelligence continues to get thrown around as a marketing buzzword from a lot of companies, it’s important to understand what it is and what it can and cannot do, especially in a factory automation setting.
In the future, it might not matter whether the approach to solve an inspection is example-based or rule-based, or even some combination of them both. For the time being, though, each approach has their inherent strengths and weaknesses and should be used accordingly.
To learn more about the differences between example- and rule-based inspections, download our free ebook: Deep Learning vs. Machine Vision.