Artificial Intelligence Primer for Meat Processors

There’s no shortage of headlines lately proclaiming that artificial intelligence (AI) is transforming the food industry. But what does this mean in practical terms? If you’re trying to make sense of AI meat processing and the future of machine learning for food industry leaders, check out this must-watch interview featuring James Spere, Head of Software and Data at P&P Optica.

James lives and breathes AI and machine learning in his day-to-day role. In this video, he covers how AI differs from automation, why data is so critical to AI success, the role of AI in detection and prevention, and how this technology is overcoming the unique challenges inherent in meat processing plants.

Time to watch: ~ 16 minutes. Don’t have time? Skip straight to the highlights!

AI in Meat Processing: Video Highlights

The difference between automation and AI, explained through the lens of packaging (2:30) :

“Automation is simply: I have a box. I’m going to lift a thing into the box. I’m going to tape the box. I’m going to ship it. AI applied to that same sort of analogy would be: I’m going to understand what I’m packaging and what time of year I’m packaging it. Did the shape change? Did the parameters change? And I’m going to change the box. I’m going to change the size. I’m going to change the tape structure. I might even change the packaging material itself based on all that. So automation is sort of the – how do I do something very repeatable over and over and over again with no brain? AI takes it to the next level of – I’m going to understand the variability and the problem.”  -James Spere, PPO’s Head Of Data and Software Development

How meat processing poses unique challenges for AI (3:53):

“The biggest challenge in meat manufacturing, and really any manufacturing, is pure variability. Whether it’s plant environment – humidity, temperature, belt speed – right down to product variability…​​as you’re trying to decide what to put in place in your plants, the most critical and key component here is a partner that understands that variability.”  -James Spere, PPO’s Head Of Data and Software Development

Why data is essential to effective AI (4:35):

“One of the amazing benefits of [PPO] and our solution is that we process millions of pounds of product a week, which means we see billions of data points per week. If you took Netflix’s entire data library – including all of their videos, all of their usage patterns, everything that’s streamed – we have seen two times the amount of data that Netflix has in their twenty-plus years of being in existence. With those billions of data points we’re making millions of decisions a second on what is essentially your product and what shouldn’t be there. So, one of the major aspects of us having that data is we have the experience to understand how those changes and how that variability is manifested in the ultimate end product.”  -James Spere, PPO’s Head Of Data and Software Development

The role of AI in detection and prevention (11:42):

 “We have a customer in the chicken industry that right now we’re detecting an awful lot of ingesta that isn’t supposed to be there. That is allowing that customer to go back to their vendors and their suppliers and really explain that their process is falling short. It keeps bad product out of the production stream…and it’s holding everybody in that entire process accountable.”  -James Spere, PPO’s Head Of Data and Software Development

Keep Learning with More Behind the Scenes Videos

Read a Transcript of James’ Interview:

Jump to a section of interest:

• AI and Machine Learning for Food Industry: Backgrounder
• The Difference Between Automaton and AI
• Challenges With AI Meat Processing
• The Importance of Data for Decision Making
• Using the Data to Learn What Normal Looks Like
• Reliability and Trust in AI
• A Comprehensive System Approach
• The Role of AI in Detection and Prevention
• Getting Started with AI Meat Processing
• The Role of People in an AI World

Heather Galt: Welcome to Behind the Scenes with PPO, a video series that looks behind the curtain at how meat processors are working with tech companies like PPO and why.

I’m Heather Galt, the Chief Customer Officer at P&P Optica. And I’m here today with James Spere, the Head of Software and Data at P&P Optica. James has spent the last twenty years working in software, including five years in AI and machine learning. He’s also fantastic at explaining complexities of AI machine learning in ways that make it easy for people like me who are nontechnical to understand.

Welcome, James.

James Spere: Thank you.

AI and Machine Learning for Food Industry: Backgrounder

HG: Let’s start with a little bit of history looking backwards. Tell me where AI and machine learning actually came from and why.

JS: Yeah. When you look at any sort of manufacturing aspect, there’s a lot of human involvement. And humans have to make a lot of decisions. They apply a lot of knowledge to their jobs, but the simple truth is humans get tired. They get sick. They get bored. They get distracted.

And so when you look at where to put in a solution, you start with automation. How do we replicate the human aspect?

The next big jump in the evolution is how do I then make that automation dynamic? How do I adjust to your changing environment, your changing needs, minute by minute, second by second? And how do we really move and change what people do to the higher value jobs for your organization, and really let the computers – that don’t get sick, don’t get tired, don’t stop making those decisions – do the things that they’re really good at doing.

HG: Got it. Okay.

Automation vs AI

HG: So, people hear the terms, and you mentioned them both: automation and AI. Are they interchangeable or do they mean different things?

JS: They do mean different things. And when you’re looking at replacing aspects of your manufacturing plant, you’re going to hear all sorts of terms. You’re going to hear automation. You’re going to hear AI. You’re going to hear machine learning and many, many others.

The simplest aspect of this is automation versus some form of AI, or artificial intelligence. If you think of an example of packaging, most manufacturing facilities will have some aspect of packaging, putting things into a box.

Automation is simply: I have a box. I’m going to lift a thing into the box. I’m going to tape the box. I’m going to ship it. 

AI applied to that same sort of analogy would be: I’m going to understand what I’m packaging and what time of year I’m packaging it. Did the shape change? Did the parameters change? And I’m going to change the box. I’m going to change the size. I’m going to change the tape structure. I might even change the packaging material itself based on all that. 

So, automation is sort of the – how do I do something very repeatable over and over and over again with no brain. AI takes it to the next level of –  I’m going to understand the variability and the problem. I’m going to try to emulate a human decision. That is the main difference between those two.

HG: So, is that why when my Amazon package arrives, it’s always in a different sized box?

JS: Sometimes. Amazon’s getting very good at changing their packaging, obviously, for the environment as well. But, you wanna receive your package safely, securely in one piece, and so adapting the packaging to it is a very key aspect to their AI approach to packaging.

HG: That’s very cool.

Challenges with AI Meat Processing

Everybody sort of understands Netflix and the streaming service. If you took Netflix’s entire data library – including all of their videos, all of their usage patterns, everything that’s streamed – [PPO has] seen two times the amount of data that Netflix has in their twenty-plus years of being in existence. 

JS: The biggest challenge in meat manufacturing, and really any manufacturing, is pure variability. Whether it’s plant environment – so humidity, temperature, belt speed – right down to product variability. And product variability can be how long it’s been since it came out of the freezer, or how long it’s been since it was post processed. If we take it right back to the farm, it can be how the animal was fed and taken care of, and what time of year it was processed.

All of those variables play into the end product that we look at. So, the biggest challenge is that variability.

One of the amazing benefits of us and our solution is that we process millions of pounds of product a week, which means we see billions of data points per week.

The Importance of Data for Decision Making

HG: Can you give me an example of what that’s meant in a plant? Are there any examples where PPO is one of a few different modalities being used?

HG: Wow.

JS: With those billions of data points, we’re making millions of decisions a second on what is essentially your product and what shouldn’t be there. So one of the major aspects of us having that data is we have the experience to understand how those changes and how that variability is manifested in the ultimate end product.

And so as you’re trying to decide what to put in place in your plants, the most critical and key component here is a partner that understands that variability. Because that variability is not only going to be different from day to day, hour to hour, but it is your proprietary information. It’s what is your product. It’s your secret sauce to producing what you produce.

And so one size doesn’t fit all here. Pairing with somebody who can adjust to your production process, your product, and your quality standards is critical in choosing what to automate and what to replace with AI.

Using the Data to Learn What Normal Looks Like

HG: That brings up a good question for me: PPO’s gathering all this data every day. You said billions of data points. How does that get added in? Like, does it automatically happen? Is it something that someone has to go in and actually work on? How do you take all those billions of data points and learn from them?

JS: Yeah. The truth with AI, as much as it sounds automatic, it’s a little bit of both. So some of it is automated learning. And, to really simplify it down, our system learns what normal looks like, and we’re able to then sort of assess anything that’s not normal. And that includes everything from low density plastics to wood to cardboard, all the way up to things like ice on your production line. This is really key, truthfully, because if you’re bringing in product from a supplier and you’re paying for product and instead you’re actually getting ice, you want to know that. And while your main concern with foreign material detection is the foreign materials, we’re able to assess every single aspect of your system.

There is also some manual effort here, and it’s really key that when you look at any AI-based system, you realize there is a bit of ongoing cost to that implementation. And that is your experts essentially telling the system when things have changed.

If it’s a false detection. If we’re thinking foreign material detection and you have something that was being rejected that shouldn’t be. Chicken skin isn’t necessarily a foreign object and you don’t want rejecting on that. You tell us, you tell the system, and the system learns and never makes that mistake again. 

So, there is a bunch of automation and automatic learning. But, like anything human or computer, teaching it is the key to it never making the mistake again.

HG: Got it. Okay. That makes a lot of sense. Thank you.

Reliability and Trust in AI

HG: So, one of the things meat processors worry about all the time is reliability. And AI has been around for a long time, but as you just talked about, it’s learning all the time. So, how does meat processor know they can count on AI, and what are some of the things they need to pay attention to to make sure they can keep counting on it?

JS: With any system – whether it’s vision-based, x-ray-based, metal-detector-based – you will have a verification plan in place. This is a standard daily, hourly, completely up-to-you process of how you validate that the system is doing what it’s supposed to do.

On a good day, if everything’s going well, you could have zero foreign material detections. Nothing’s been introduced up the line. You don’t have a plastic problem. You don’t have a wood problem. And the hardest thing to sort of trust is: do I not have a problem, or is the system not working?

And if we come back to the PPO system, we have built-in tests through a calibration. So we understand if the system is in alignment, in spec. And we provide verification targets that essentially mimic real-world tests that you can run whenever you need to throughout the day. So, it is really key that you’re ensuring that this system is working.

This extends to, again, x-rays, metal detectors, same sort of aspect with a verification bar. The other aspect is when you partner with somebody. Again, I’ll come back to PPO because that’s near and dear to our hearts. You have a team on the ground with your provider. And, so, at PPO we are constantly monitoring in almost real time the trends of your system. 

So we will know, and you will know also as an aspect of our reporting structure…Are things changing? Am I seeing spikes in certain materials?  Am I seeing spikes on certain days? 

And those are also important aspects because it isn’t just at the point of your foreign object detection. It is a whole process. 

So, if you can start tying back to Shift A on Mondays seems to have a little bit more of a problem than Shift B, it helps you train and educate, and change your processes. 

So you’ve got to think of this as an entire system. Your ultimate goal is to deliver the highest quality product, and our ultimate goal is to help you deliver the highest quality product. But that doesn’t just happen at the spot of detection. It happens all the way back from when the product comes into your building, your processing, your cooking, your post-cooking, and at the spot of detection. And so you have to be able to understand how all that ties together, and, luckily, PPO does it all for you.

A Comprehensive System Approach

HG: It sounds like our system, as you’re as you’re describing it, is almost like one piece in a bigger system of not only detection, but also prevention. And sort of monitoring how the whole thing is working is one of the best ways to make sure it continues to operate as you expect. Is that correct? 

JS: Absolutely.

The Role of AI in Detection and Prevention

JS: And there are aspects of – again, I used the ice, example – but we have another customer in the chicken industry that right now we’re detecting an awful lot of ingesta that isn’t supposed to be there.

That is allowing that customer to go back to their vendors and their suppliers and really explain that their process is falling short.

It keeps bad product out of the production stream. And in that case, we’re very early in their production stream, so the product that’s not supposed to be there doesn’t get chunked up, doesn’t get mixed in, doesn’t get shipped.

They’re able to go back to that supplier and  say, “Fix it, or we’re not ordering from you anymore.” It’s saving money, it’s saving time, and it’s holding everybody in that entire process accountable.

HG: Right. Not just in that plant.

JS: Exactly.

HG: Yeah. That’s awesome. 

Getting Started with AI Meat Processing

HG: That brings up a really good question because if you’re worried about AI, you know, clearly it’s delivering exactly what meat processors need right now.

So how can a processor get started? Because AI is not natural for any industrial application. 

JS: Yeah. It’s not. And with the meat industry…we talked about the complexities, the variability.

If I’m sort of looking at my plant operations, I’m looking at a problem that I need to solve. And in many of our customers, all of our customers, foreign materials are an issue. Mainly because the current solution is humans. And humans, by nature, we’re actually programmed to not do repeated tasks very well. Our minds drift. Again, we get tired. We get sick. 

So, when looking at that, our customers have sort of two problems. One, humans aren’t perfect. And two, there is a general labor shortage in most of the manufacturing industry, meat processing included. 

You want to maximize your human capital. So where is the better place to put those people? It might be in packaging. It might be in preprocessing. It isn’t necessarily in foreign material detection. So you want to look at a problem that you have, and you want to see what solutions are out there.

No detection system is going to catch everything. It is part of a multi-hurdle approach. You might still have some humans there. Our system is very, very good at detecting anything it can see. X-rays obviously detect within, and so you want to look at what your multi-hurdle approach is to foreign material detection, but understand the problem you’re trying to solve, understand your own success criteria, and then work with a partner that can actually deliver on that promise.

HG: Got it. So, really, I think one of the keys to success then is actually understanding what you’re trying to do. 

JS: Absolutely.

The Role of People in an AI World

HG: Okay. We’ve kind of gone deep into the meat processing industry, but let’s pull back out.

So, what do you see being the role of people in an AI world? It’s a big question for everybody. I’m curious to hear your answer.

JS: Yeah. Again, AI is not this big, bad, evil thing that’s trying to take over the world and replace people. People are not great at making the same decision over and over and over again. And so AI in all of manufacturing is a very key contributing factor to automating the repeatable, the sort of mundane jobs that you and I wouldn’t necessarily want do it every day either.

And so it allows you, again, to take that human capital and distribute it differently in the plants, whether that’s meat manufacturing…we talked a bit about Amazon’s packaging…anything that gets manufactured has some human aspect that can absolutely be replaced and repeated by computers.

Whether it’s meat or production, again, I’m looking at where do I have my human capital now, and what solutions are out there that can help me realign those and redistribute them? And then people become both the teachers (how do we provide that feedback loop?) and then people move into other more valuable spots in your organization.

HG: Where maybe they’re gonna stay a little bit longer?

JS: Absolutely. 

HG: Awesome. Okay. So it sounds like I probably don’t need to worry about losing my job to a computer anytime soon.

JS: Not yet.

HG: Okay. I was a little bit worried. Alright. Thank you, James! That was awesome. I really appreciate that.

JS: Anytime.

HG: Don’t forget to check out our other Behind the Scenes videos, including our interview with Philip Hancock, where he showcases some of the amazing things the PPO system has been able to find using our patented hyperspectral imaging system and artificial intelligence. See you next time!