The Future of Manufacturing passes through AI

Cleaned up from the automatic captions and shortened. Cuts are marked […], and where the recording is unclear the wording is my best reconstruction. For the exact words, watch the video.

Host: Our next guest is connected from Boston: Massimiliano Versace, co-founder and CEO of Neurala and founding director of the Boston University Neuromorphics Lab. […] Would you briefly introduce Neurala?

Max Versace: A pleasure to be here, and to be virtually in Italy again with my friends. I started the company with Heather and Anatoly, my colleagues at Boston University, in 2006. It grew out of the research we did there, with NASA among others. Many years later, as the industry matured, we took this technology down from Mars, where it was supposed to run, to more terrestrial applications. […] I will tell you about the journey from designing things for Mars all the way down to production lines.

When people think of AI and robotics, the first things that come to mind are robotic kangaroos, dinosaurs and snakes. That is misleading. The real incarnations of AI look very different from the Hollywood images: agriculture, automotive, retail, healthcare, and of course industry.

What about manufacturing? Deep learning has the lion’s share of the growth, and within deep learning, the part that lives on the device, at the edge or on premises, is the one that matters most for what is feasible in a factory. Then 2020 added a twist, the coronavirus. The pressure on manufacturers is even higher: increase efficiency, maintain and improve quality, reduce waste, keep lines flexible. […]

Where are we on the road to Industry 4.0? Roughly at stage three. After a good amount of digitization we can collect data, so we have visibility. Stage four is: now that I have all this data, what is it telling me? Collecting data is not a panacea. You end up with a large bucket of data. The real step is from visibility to understanding, and the way to do that is with AI, because no human operator can sift through that amount of data.

What is this AI about? We know much more than twenty years ago, when I entered the field. We can describe mathematically the basic working of a brain, human or animal, and translate it into software. It is very far from our own capability, but good enough to be useful in the real world.

There is a prehistory of AI, the so-called good old-fashioned AI, and there are neural networks, which is the real name of deep learning. In good old-fashioned AI the scientist breaks the problem into explicit steps. To tell an apple from an orange you build a decision tree: is it a fruit, is it roundish, is it green? In a neural network the intelligence of the scientist goes into the mathematical rules that connect neurons and synapses, and it stops there. Then you present data, the way people learn throughout their life, and the network changes its structure to tell the apple from the orange.

Anybody who has babies knows that most of the learning happens without supervision, just by being exposed to their environment. I wanted to teach mine something, and I failed.

[…] Neural networks started to win the academic competitions, in image recognition, language, traffic signs, medical imaging, and then they started to flood the real world. Over the years we have put our AI in phones, where it works like a professional photographer each time you take a picture; in drones that inspect a wind turbine or a power line so that you do not have to send a person up there; and in hundreds of robots that roam grocery stores.

In manufacturing, AI is already used to check that the expected items are in a box, to inspect labels on personalized packaging, to check that pallets arrive in good condition. But AI and deep learning are not miracle technologies. They are technologies, with very specific operating conditions and, if you will, defects.

Imagine a bottling facility where you need to inspect caps. The traditional way to build a deep learning system is to collect thousands of images for each type of cap, with all the anomalies that can occur. And you need what the jargon calls a balanced data set: as many bad examples as good ones. Then you train for hours or days, and once it is trained it is fixed. To change it you go back to step one.

The balanced data set is the big issue. Fortunately, the number of defects that come out of an industrial machine is really small, so you never have half of your images showing a defect. So how do you do it? The brain comes to the rescue. For us it is innate to know what normal looks like. Show me the Mona Lisa with something changed and I spot the anomaly at once, with no effort, because we continuously match our expectation of normal against what comes to our senses.

That is what we built into our technology. For manufacturing it changes the paradigm. You collect a number of good images, which is what the line produces 99.9% of the time, you train on good products only, and you deploy. Whenever something deviates from that golden standard, the AI alerts you. You can do it in minutes, without knowing anything about the AI underneath. […]

That is vision. But an industrial machine has a plethora of other sensors, and their data is already being collected. Think of the cost of unexpected downtime, when something breaks. We can use the data from the machine before it goes down, to be strategic about when we take it out of service.

This is where our collaboration with IMA comes in. The first application is on a coffee capsule machine, where a sensor checks the quality of the seal at each cycle. As with vision, we feed the AI the sensor data from good production cycles, we train on what a good cycle is, and then we use it to spot the anomaly. It is a step towards machines that are more aware of what is good and what is bad, and that can alert the operator before something goes wrong.

To understand the future it helps to look at our past. We worked on artificial brains for machines like a Mars rover, which has many sensors and has to adapt in real time to an uncharted place. There is a strong similarity between an autonomous robot and an industrial machine. The machine has no wheels, but it has an equally impressive set of sensors: vacuum, temperature, motor force, pressure, seal checks. Together they give deep insight into what the machine is doing, and into what is happening when one of them goes astray.

The trick is that each machine is different from its cousin: a different footprint, different sensors, different products. So each machine needs its own brain, the way a rover needed its own, trained on what normal is for that machine and that production run.

Collecting the data takes you halfway around the circle. What closes the circle is the AI that integrates the data and understands what is happening in the machine. We are going to do more with less, because the sensors are already there. Thank you, and my apologies for the connection.

Question from a viewer: Have you worked with other users in manufacturing besides IMA?

Versace: Yes. A few weeks ago we launched our visual inspection product and a partner network that distributes the technology on virtually every continent. […] The adoption we are seeing is very fast, and I think this year has created a sense of urgency among manufacturers all over the world.