AI was inescapable in 2024. What once was a science fiction acronym became a buzzword in almost every industry. From aerospace to agriculture, banking to biotechnology, seemingly every sector was scrabbling to think of ways to maximize their use of machine learning.
In the food and beverage industry, this brainstorming led to proposals of using AI to automate food production, streamline supply chains and personalize consumer nutrition. But is any of this feasible? Are the new, hyped forms of generative AI even applicable to food production?
To learn more, Technology Networks spoke with Nicholas Watson, a professor of AI at the University of Leeds.
Whenever you talk about AI, you have to think what is AI? So, I’d always say to someone, “What do you think AI is?”
Well, exactly. AI has been around in our computers for many years, but it’s suddenly become thought of as something else in the dawn of chat bots like ChatGPT.
Yeah. And I think this is the way to think about it; most of the time when people say AI, they mean machine learning, which is where you have a problem and you’ll collect some data, then you’ll train models to make predictions. But the worrying thing these days is that people use ChatGPT and they think that is just what AI is, and that’s a generative AI; it’s a large language model. But there are many other different types. AI is the ability to make decisions – identify patterns by non-biological entities. That’s what I think AI is. The way I always describe it to people is it’s about making better decisions faster and cheaper.
In terms of food, I think there’s lots of areas you can use it in to make things more efficiently. So, for example, we might want to reformulate a food product to make it have a better environmental profile, a better cost or a better nutritional profile, or, ideally, all of those. Or it might be that we’re trying to take some waste, such as food waste, and upcycle it into something that people can eat as food. These things just take a lot of time to do. There’s lots of trial and error, lots of experiments. So, the way we see AI is, if we can collect some data and build models, we can actually reduce the amount of experiments we need to do and reduce the cost. So, again, it’s enabling us to make better decisions faster.
I can think of a few examples. One of them is a project we’re doing with partners in the UK and in Australia about food waste. Typically, if you want to reuse food waste, you can turn it into a liquid media, then you could ferment that with yeast. Then the yeast will procreate and you’ll get a microbial protein, which you can then process into food. But there are lots and lots of different parameters. How do you treat the waste initially? How do you understand its varying composition? And then you go to fermentation and you’ve got things like temperature, the type of organisms you use, pH, how long you ferment it for. This is just a big problem. So, with AI, we can actually take the number of experiments and reduce the time and cost, from maybe 25 experiments to about 5 experiments. We use some of our own data, but then we go to the literature and use tools to extract all the information and data in that literature to augment what we’re doing. That’s one example.
You asked what is AI? So perhaps it’s good to be specific and clarify whether the AI phenomenon that everyone has been interested in for the past year has really changed things in the food sector. Or has machine learning been integrated into research and production for a while?
So, the answer would be yes and no. I mean the models we’re using have been around for decades. Artificial neural networks are the main one. There was a big boom around 2005; researchers in Toronto figured out a way to train these models. Around the same time, you had the rise of the Internet of Things, more devices connected to the Internet. So, we’ve suddenly got much more data coming in to train the models. And you’ve also got increased computing power. So, most people use cloud computing now on high power.
The large language modeling is interesting. We use it if we want to go into the literature and find lots of information. If I just sent you 100 papers and said, “Right, read these, pull out all the data,” that’s going to take you months. Now you can use these things, and they will do that, extracting that information in a matter of minutes. But it brings challenges in its own way as well. What if it brings in the wrong information? What if it misrepresents things? Or what if some of those papers in the first place had errors in them? Everyone’s human; people publish stuff with mistakes in them, so that then goes into the model that we’re using to make decisions.
One of the challenges in doing research in general today with these enormous models is you need massive computing power, which most universities don’t have. It’s with the big companies. You also need the kind of training data sets to benchmark against, which, again, we don’t have. I’ve read some interesting papers in the last year about how it’s virtually impossible to be an AI researcher unless you’re working at one of the powerhouse companies, or working in collaboration with them.
Interesting. I’ve had to correct Gemini a couple of times myself. Do you think there is a good robustness in these AI language models, then? Do you trust them yet?
It’s a really good question. Because one of my bugbears is that the expectation we put on technology is often a lot higher than the expectations we put on humans. The classic example is self-driving cars. If you look at any statistics, self-driving cars are way safer than human-driven cars. Yet every time there’s an accident with a self-driving car, it’s on the front page of a newspaper.
Context is really important as well. So, if we’re using AI for inspecting the quality of food – let’s imagine you’re a food factory and you’re making a wonderful product – you’re probably taking a sample off the production line every hour. And a human is looking at those and going, “Does it meet our quality requirements?” Well, first of all, you might not spot the problem if it’s just a random one; you might just pick a sample off that doesn’t have it, so you won’t find out until people start complaining. Or it could be that you’ve had this error for an hour and you’ve produced an hour’s worth of food, which you’re either going to have to rework or waste. If you put a sensor AI system in there, you could probably image every single food in real time. Now, does that really need to be that accurate? Probably not, because the fact that you’re measuring every single one, if there is a problem, even if it’s low accuracy, you’re probably going to pick that up in the trend and you can do something about that right away. So, for that, you don’t actually need it to do it that accurately or perfectly. The fact that it can do every single one is just great.
The other side of that is food safety. One of my PhD students has been working all week to get his model accuracy from 99.5% to 99.55%, and I’m like, “What does that actually mean in terms of the problem you’re solving?” Because if we’re looking at, say, allergens in food, 99% accuracy – which is generally good in most training models – means 1 in 100 is wrong. And if you’re making hundreds of thousands, if not millions, of products a day, that’s a lot of potential risk you’re making. So, I think it’s about how do we manage that? Maybe we have this AI system as an early warning indicator, but we still continue with the very robust food safety protocols we have taking samples and sending them off to the lab.
Often the mistakes with AI are what we call edge cases. So, if you’re making pizzas and the pizza base is all black and burnt, well, it’s not going to get that wrong. But it might miss something at the edge.
I actually think some of the opportunities are to connect data sources across the whole value chain to make better decisions. So, if you’re taking complaint data from consumers and feeding it back in real-time to the factory, you can be making the right quality for the people who are going to have it. And that’s really interesting.
And when you speak to the folks in industry, is that what they’re keen on, that kind of interconnectivity?
Yeah, absolutely. I think they see it as a unique opportunity. There’s lots of people doing AI in areas of food, and there’s some really great work going on, but it’s often in silos. That’s the interesting thing about working in AI. Let me ask you, why are all these big companies AI companies? What have they got that no one else has?
For the processing power.
Yeah. What else have they got that’s even more important?
Money.
What else is even more important than money for AI? We eat food. What does AI eat?
Exactly. They’ve got these large datasets that they can use, and it is your data, it’s my data, it’s personal data. And some of the companies we work with make millions of their products a day – that’s millions of samples for our training data set. So, we need to work with industry partners. They’ve got the data; we’ve got ideas and know-how.
It’s interesting how both AI and AI projects seem to thrive on greater connectivity. Like you say, you have the know-how, they have the processing power.
I mean, the processing power is really only a problem for these super massive, large language models. A lot of the work we’ve done is just on general CPUs [central processing units].
People always say to me, “How much data do you need?” You need enough data to represent the variability in the system you’re trying to model. So, if we’re looking at qualities of pizza, we need images or other sensor data and all the different qualities of pizza and all the variability we can expect, related to the base, the toppings, to the sauce; we need that variability to train the data set. That’s what’s absolutely key.
It’s not just about collecting the data. You can imagine it’s almost trivial to put a camera in a pizza factory and record every pizza, but a lot of machine learning uses a technique called supervised machine learning, so you have lots of data, and then you label that data. So, for the pizza example, we would have good quality, bad quality or we could have a quality score from 0 to 10. Now, capturing a million images of a million pizzas is just a case of putting a camera there and leaving it. But someone has to look at every image and say, “Good quality, bad quality, 0, 3, 4, 8, 10…” and that labeling of data is often what takes the time. So, how do we overcome that data labeling burden? We’ve used methods such as transfer learning, active learning, semi-supervised learning, where we label some of those edge cases which are quite hard to determine.
So, this level of subjectivity is where the human needs to come in and modulate?
Yes. We’re doing a lot of work on alternative proteins with our partners at Sheffield University, Imperial College London, James Sutton Institute and loads of others; we’ve got over 100 industry partners and about another 20 academic partners.
One of the challenges you have with plant-based proteins is the bitterness and the taste. They’re just not that enjoyable to eat. Some people might like that, some maybe don’t. So, we’ve got a research fellow who just started on a very prestigious scholarship, and his whole work is about how you can select and process different proteins to reduce that bitterness level. That’s what he’s really into, that stringency level. So, he will collect some data, he will try different types of plant-based protein and different processing techniques, and he will measure their bitterness; then we can build a model that will say, “OK, for this protein and this processing route you’ll get to this bitterness level.” But what we actually want to do is try and run some optimization on that, because we don’t want to predict bitterness; we want to reduce it. That’s generally what most people want. But we could also use that to change it for flavors. I just came back from a few weeks in China, visiting the university for a conference, and we had the pleasure of eating something called stinky tofu. Have you ever had stinky tofu?
When you get off this call, just go on Wikipedia and put in stinky tofu and have a look at the flavor. It’s one of the most interesting things I’ve ever eaten. I straight away spoke to some of my Chinese friends and said, “How can you possibly eat that?” And they went, “What are you talking about? It’s lovely. That blue cheese stuff you guys eat, what is that all about? That just tastes like sweaty socks.” It’s just that cultural thing.
That again, is something you can do with AI; you can use clustering methods to identify what is someone likely to like. So, imagine you walk into a restaurant and there’s facial recognition software. It’s scanned the microchip, and you put your loyalty card in, and it goes, “OK, Leo’s always buying these kinds of things, and he likes this kind of thing. So, let’s cluster them.” Much like how recommender systems work on Netflix and other streaming patterns; you like these programs, so it’s going to recommend ones other people like you like; it’s going to put you in a cluster. It could say, “Right, we think when he orders his curry, he’s going to want this spiciness level. So we know we can do that.” We could do all that autonomously without even telling you. But then there’s the human bit. Some days, maybe you like a mild curry; other days maybe you want a really spicy one. For me, these are the interesting questions we need to explore in AI; it’s the human part of it, the intelligence part.
I mean, do we actually understand what intelligence is? If you ask Gemini, it will give you 10 definitions of AI. And it’ll use the word intelligence like we really understand what that is. I think these bits of things are what we should spend our time exploring.
