AI adoption in large, well-established companies

Welcome to another episode of Practical AI. This is Daniel Whitenack. I’m a data scientist with SIL International. We’ve got a very special episode today. Way back at episode 120 we talked with a couple representatives from the company Aryballe, and we talked with them about this very interesting subject of how AI and sensor data is transforming this industry called digital olfaction, which is using advanced AI technology to actually create fingerprints of odors or fragrances, and use that to transform industries like manufacturing, or the fragrance industry, or the food industry… And I got a chance to follow up with Yanis from Aryballe and some others at a digital olfaction summit, and we had a follow-up discussion about how AI technology is transforming organizations, and how organizations are responding and implementing this sort of technology. It was quite interesting in terms of the adoption of AI at a large scale, and how that is actually shifting large organizations and how they think about the problems that they’re solving. So I hope you enjoy this panel as much as I did. Here’s the recording from the summit.

Thank you. Yeah, it’s really wonderful to be here at this super-exciting summit. I just looked at the date, it was January of 2021 on the Practical AI podcast, we had a topic focused on how artificial intelligence is transforming this space of digital olfaction and detecting scents and fragrances and all of that, and it’s awesome to kind of come full circle and see maybe a follow-up of how artificial intelligence is transforming this industry and making an impact for a variety of companies.

I’m excited today to have on the panel with me Mary Fischer-Mullins, who is a Senior Director of Project strategy at Cox Automotive, and also Yanis Caritu, who is the chief software officer at Aryballe. Welcome to you both. Glad to have you on the panel.

Thank you.

Thank you, exciting to be here.

[03:54] Yeah. To start with, I thought generally maybe we could start talking about from both of your perspectives how artificial intelligence and the use of artificial intelligence within your organizations has changed over the past five years. So that could be in relation to this area of scent, fragrance, digital olfaction… But maybe more generally than that, how perception of AI has changed within your organization and adoption of these advanced technologies has changed. Maybe I’ll start with Yanis, because we are at this summit… From your perspective, and – you’ve been working to develop these technologies specifically for digital olfaction over the past years… How have you seen specifically artificial intelligence and that side of advanced technology be a key technology within what you’re trying to build?

Yes, thank you, Daniel. And we can say since the past five years.

Yeah.

So yes, we talk from the production and R&D point of view maybe. At Aryballe we use machine learning instead of artificial intelligence. You know, there is an endless debate around this terminology. Machine learning is very well adapted to our problems, and the amount of data available… And as you know, machine learning is really mandatory when designing odor sensors. So for us, the question five years ago was not if we will use AI, it was how we will use it. And that being said, we use it with discernment. AI has promises, that’s true, but it can be a source of disappointment sometimes, if you don’t take some precaution. And that’s what we experienced also.

So what changed is that we need to grow with experts in digital processing, and to avoid the traps of, I would say, AI [unintelligible 00:05:53.24] We grew a team with data scientists, but also with people having a solid background in physics, chemistry, or signal processing. And this is our DNA, to think about best ways to use machine learning in the sensors area, but also, as a consequence, we spent much more time on understanding our data and pre-processing them more than on machine learning pipelines and cells, which are now commodities.

So I think also in our use of AI we also became more professional in using this technology. We extended our capabilities by using the cloud, of course, for hosting all our stuffs, and measuring databases for our customers, but also for us… Knowledge and contextual databases, which are very important in AI, and so on. Also internally, we built our own data center; maybe “data center” is a big word there, but it was that, and also, to facilitate and accelerate the work of using artificial intelligence for our R&D and product integration, I would say.

Thanks, Yanis. What about yourself, Mary, in terms of Cox Automotive and how maybe the mindset and strategy related to AI-enabled technologies has shifted over the last five years or so?

Absolutely, yeah. AI means a lot of things to a lot of different industries, and if you had five people – if you had all of us in this room and asked us for the definition, it would all be different, because we’re all using it differently, in some sort or fashion. And to Yanis’ point, sometimes it can bring about disappointment. So it’s important to understand what you’re looking at, and how you’re approaching it. I think when I go into research, or I take on a new topic, one of the things that’s important is to have that open mindset to say “I think I know what I’m going to discover, but I need to look for the things that kind of bubble up along the way as well”, the unexpected exciters or disappointments that you may find along the way. And I think that’s available across all the different spaces.

[08:08] We are using technology every day. I’ve been working on a specific project to really understand imaging technology, and we started that well before the pandemic. And as everybody knows, on one Friday, our business was normal around the world, and mostly in the US come Monday, that fateful Monday, one day in March of 2020, our businesses shifted, and we had to – and we really shifted into more digital platforms, much quicker and much faster than we anticipated. And that kind of really progressed the way we’re approaching things and looking at things to understand how do we provide a quality, repeatable experience for our clients and our partners in the ecosystem, and AI is really helping us do some of that. It helps us to build that trust and competence in our products, and in the experience… And so we’re really working to understand what else is out there, because technology is changing so much every day.

Yeah, that’s super-helpful to think about that… And also, to think about how just the global shifts over the past few years have forced a lot of people to maybe consider things on a timescale that’s different from their long-term roadmap.

So next, I want to talk a little bit about like the next opportunity that’s on your mind in terms of augmented or AI sort of advanced technology within your organization. Maybe this time I’ll start with Mary… I know you talked about imaging, but also, we’ve talked a lot at this summit about the sort of fragrance and scent and advanced technology in that space. Where are you thinking in terms of the next challenges that advanced technology might – where that might be most applicable in your organization?

Absolutely. So when you talk about that transformation, that quick transformation of our roadmap, as just our journey, whether we’re focused on one thing, or a bigger company, we looked at this and said there’s emerging technology when it comes to odor-sensing tech, and how can we leverage that, or just explore it to see if it’s… You know, I kind of look at it from the scope of, “Is it ready now? Does it apply to our potential scenarios that we have in our ecosystem? And is it repeatable? It goes back to that trust and competence. Can we repeat and create a consistent and valued experience out of the transaction itself? And when we look at the digital world, one of the things that we’re working in the automotive space is to say – you know, we touch three out of four cars on the road, and we have all these varied ecosystems, whether it’s mobility, or whether it’s dealer-to-dealer, or consumer, or partnering with our OEMs… You know, we’re looking to understand how we can present better-quality information to them in a digital experience, and we think that odor has some potential to explore and understand that.

[11:48] It’s interesting to kind of say – you know, if you’ve ever bought a pre-owned vehicle, and perhaps it was something that maybe was used as in an industrial application, or with somebody who maybe smoked in their car… Some people are more sensitive to those kinds of experiences, and we want to create a value that we can express consistently in our ecosystem, that then allows us to create that output that gives our client portfolios what they’re looking for. Imagine if you’re back in March of 2020, you suddenly had to transact vehicles digitally, and you couldn’t be there to touch, and feel, and listen, and smell the inside of the car. And if you’ve ever talked to buyers and sellers, they want to stick their head in the car, and just give it a little sniff. Because as we all know, scent, and looking at the presentation and listening to everybody today, I kind of thought, you know, scent is really a language that we all speak. Our memories are grounded in it, and our brain triggers very specific things for us… And I think that language is important to us all in our different spaces.

We spend so much time in our transportation modes, whether it’s a vehicle, maybe it’s mass transit- maybe it’s monitoring inside of that space, but we want to understand what makes it comfortable and what makes it create value, and perhaps what excites us about that scent in our experience with transportation.

Yeah, it’s interesting you bring in a sort of language parallel… Most of my work day to day is in natural language processing, and I think there we have the advantage that everybody knows about language, but there are common ways to represent it in letters, for example. Whereas scent, odor, it’s sort of something we are also all familiar with, but having the mechanism to represent it objectively is a difficult thing, versus like subjective terms. So it’s interesting to hear how the ability to take odor with an AI-enabled platform, and kind of process that off of sensors, and make the odor sort of signature or fingerprint something that’s tangible creates a kind of language that allows you to do various things. So that’s really interesting.

Yanis, I know that you’ve been instrumental in that process of taking this sort of fuzzy, intangible thing, odor, and kind of using an AI-enabled system to make that more tangible via the Aryballe system, and the platform, the fragrance signatures, and fingerprints… As you look to kind of the next challenges of the system you’re developing, what do you think are the big challenges or opportunities that AI might address within what you want to do next as a company?

Yeah, I will talk about two next opportunities we have, and we can talk about challenges afterwards maybe… I think the first is [unintelligible 00:15:01.27] with our application. For example, in the field of automotive, and some presented it some minutes ago, within the framework of our digital olfaction for automotive construction, we already have found a real study, it was last year, on a two-weeks database in real cars. [unintelligible 00:15:21.21] detection in cars was achieved with a 94% accuracy, starting from this point, to get to this odor signature, or footprint, as you said.

So this is really a successful story, but we are also building and will continue to build a large database, and you know, acquiring a database of odors is quite time-consuming, but you need to go through. With our customers, for instance, for flavors and fragrances, it is several hundred, or even thousands in the next year, of smell to collect… Of course, with automation, and there is a lot of digital tools to prepare before. Then we use machine learning to associate those measurements of unknown odors with chemical properties, or ultimately with sensations and [unintelligible 00:16:14.08] that they could create, given what we have already learned, of course.

[16:20] And the same goes for end user applications. That’s our goal for each one its own universal smell. And that’s also a discovery, that we don’t need to have a very large spectrum of odor. We need to have a very pragmatic application, and so it reduce the [unintelligible 00:16:36.28] you need to apply your machine learning.

So first, a campaign for initiating the database, then a study to analyze and learn the model, so that you can predict what the new measurement will be related to.

And the second opportunity maybe is more technical. You know, it’s very common when designing a sensor, a new sensor especially, and whatever the sensor, that you want to know how to calculate in advance what your sensor is supposed to respond when submitted to non-physical stimulus. It’s a straightforward function you want to have. With a chemical sensor, it’s quite hard to achieve this prediction. I cannot say much more today, but AI is also helping us on this topic. And as you imagine, it’s very important to have this tool to predict the output of your system for odor sensing before it is measuring something real. So it is the next opportunities, but also a challenge, to answer your question.

Yeah, definitely. I know, just from personal experience, one of the things that I’ve learned over time is that within an organization, as you’re trying to kind of roll out advanced technologies, a lot of the things that you have to plan for and deal with are more people-related than they are technology-related. Technology problems can be solved often, infrastructure things can be solved, but people’s perception of advanced technologies and how they adapt them into their workflows, and sort of the foundational/organizational changes that you need to implement as you’re rolling out advanced technologies might be the difficult things to think about.

So Mary, I’m curious, I’ll start with you… As you’ve worked with your organization over the past years to implement some of these advanced technologies, what have you found are those kind of foundational, organizational changes, or what has the experience been like in rolling out these advanced technologies, and maybe the things that you’ve hit along the way that you might be interested to share?

Sure thing. You know, when we talk about AI or advancing technology in our space - you know, go back to my imaging technology… When we think about the mobile phone over the last just four years, if you’re in the States, you may have had, I don’t know, an iPhone 8 maybe, and we’re already on the 14… And iPhone 12 and newer has LIDAR imaging technology. So it has the ability to image things differently. The iPhone 14 has some newer features. So when we talk about this simple construct of imaging itself, the technology has moved very quickly. And so when we’re building our portfolio of information and the way we’re curating that has really – we’ve kind of had some takeaways, and then we’re like, “Oh, here’s this new thing. Should we pursue it?”

So we’ve had to kind of continue to curate and understand that when we’re in the space of technology, we need to be able to accommodate across all the different formats. So it could be we need to version back to the iPhone 8 when we talk about product [unintelligible 00:19:57.14] we need to version up to the iPhone 14 and beyond. And we think about that when we also talk about our user experience of “What is the end user or the person who’s interacting with this technology, at all the different tiers, how are they going to receive this?”

[20:17] It’s interesting when you talk to a novice about artificial intelligence, sometimes they will immediately say, friends or family, “Oh, AI is going to take over, and that really concerns me.” And I kind of always say, “Well, AI is only as smart as we want it to be. And if we stop feeding it, or stop giving it quality information, it really just truncates and doesn’t go anywhere.” So we’re very cognizant of the need to continue to curate quality information, whether it’s images, whether it’s data and how we collect that, making sure that we have all the information, so that we can build this robust data lake that takes us from computer vision to machine learning and into our different types of modeling that we’re using.

So I think technology advances are amazing, and it’s almost hard in some ways to keep up with them, because they are moving so quickly. So I think in some ways we’re chasing, in some ways we’re leading, but I think those are part of the adventure and the journey and the process.

Yeah, I’m so glad that you brought up the data side of things, which… You know, there’s kind of the AI and machine learning modeling side of things, there’s the organizational changes that need to happen to accommodate that… And then there’s also like the data and the infrastructure things that go along with that, that absent that really quality information and the curation of that over time, any changes you make in the other two categories might be all for naught. So yeah, I think that’s a really, really good point to make.

Yanis, as you’ve been working maybe with various clients across different industries, have you seen any commonalities in terms of the organizational mindset, or foundational changes that organizations are having to make, now that we’ve moved from specifically your platform, enabling them to analyze odor and fragrance on a much deeper level, much more objective level? What are those sort of organizational or foundational changes you’ve seen in common across some of the organizations that you’ve worked with, that they’ve needed to make now that they have this fundamentally, or this transformation, digital transformation of something that was maybe more subjective before?

Yes, that’s a good question, but also, I think that the progress of the digitalization of everything, and in the large companies, it’s very helpful for us because it’s exactly what we should be, and it helps us to disseminate the technology easier, with the cloud, or with the many tools we are offering to our partners.

In our company, we made a big transformation, in fact, from a high-skilled R&D startup company, I would say, to a product company, still improving, of course… But that’s the most striking one for me. Even considering advanced technologies, people at Aryballe are very used to, and this is a population that really progressed quickly on this topic.

Even in the way people are collaborating, we integrated, for instance, automation at all levels, from robots for [unintelligible 00:23:35.10] or robots for sniffing odors to get more data, and also to capturing knowledge at each steps of our product chain, or even in our product. And that was mandatory.

So we digitalized odors, for sure, but we need also to digitalize as much as we could all our information, in fact. And to summarize, we try to have more and more traceability, as we know, beyond this. AI will help us a lot, if we have the data, and a lot of data, it will be very easy for us.

Maybe just a couple more questions here for both of you… What’s maybe one challenge that you see within your industry in terms of the adoption of AI or augmented types of technology? And then maybe what’s something you’re excited about as you look forward to the possibilities that this sort of technology might open up in your industry? We can start with you, Mary.

Some of the challenges - I think the biggest challenge that we all have, and it probably crosses every industry, is the human factor. Because when you put humans in this equation, which we’re all doing the work - that is our number one source of bias. So when we look at this and we talk about trust and consistency, and we want to create an ecosystem that has a quality experience every time, or quality outputs every time, we have to continually look at how we’re bringing all these elements together so that we are checking ourselves, and eliminating not just our own bias, and our own admiration for potential goals… Like, we have this lofty goal of what we potentially want to achieve, and we need to ensure that we’re setting those standards and bias aside.

The other challenge when it comes to humans is ensuring that we’re giving – you know, in our organization we’re looking to cross many skill levels and many knowledge levels, but we want to deliver outputs and tools and products and information that are consumable, that don’t feel overwhelming, that don’t feel intimidating… And so that is kind of our first goal, is to say, “What can we do that creates the value?”

What am I excited about? This has been an exciting day for me, to see how all of this crosses so many different industries. I’m excited how technology really in a positive way can change our experiences together. Technology has brought us together today, and I think those adventures of kind of the human experience and how technology interweaves us all together is part of the exciting things. I don’t have simply just one thing that I’m looking at and saying, “Wow, I wish this was happening.” I think I just take it all in and look at it with some amazement, and then I reach out to all the smart folks in our organization and around the globe that say “This is how we’re going to use it.” So it’s good stuff.

Yeah. What about yourself, Yanis?

Yes, of course, there are some challenges people in our industry face, especially with sensor, with data, amount of data, digital olfaction, and so on… So sorry about that, my answer may be probably a bit on the technical side, but I will try to be clear on specifically on digital olfaction.

When we wish to have a technical answer to a problem, we like to represent it as the information, as an input, then you have a black box, and the desired answer, which is really appealing and attractive at the output. It’s always like that. But the question is what we are doing in the black box, and it’s very attractive to put directly inside only an automated learning-based system. It may work sometimes, but sometimes only.

[27:47] So if you want to know – let’s take an example… A camera lens - it should accommodate to get an image on your sensor. There are some explicit objects below to give you this sensor, without the need for learning. And when there are no formulas, machine learning from data is used. In Aryballe sensors there are two parts: an explicit part, which we know how to describe [unintelligible 00:28:09.17] an implicit part for which we are obliged to learn from many measurements, and a certain amount, of course, of prior knowledge attached to the data. And what is interesting, the less explicit you’re making your system, the more dimensions you need to observe for the implicit part of it. And over time, we understand in our jobs as sensor guys that the use of AI to solve these problems is this mix between clearing the ground as much as possible with explicit mappings, in advance, and collecting data to feed the implicit mapping.

For example, to be more concrete, I gave this example seated a couple of times during the presentation, and it’s unpopular I think… All odor sensors are really corrupted by humidity variation during the acquisition. So we use explicit techniques with a humidity sensor, or with the amplifier, which is a hardware, to reduce this impacting the signal, and to achieve what we can call the dry smell inputs. Then we use machine learning on the remaining signal techniques to recognize the smell, for example. And if we don’t approach AI with this mix, I believe we ask the machine learning to support all the program as a black box, and we have to acquire data to fill a tube in 3D rather than a square, if I oversimplify it… And that for sure requires much more time. So the required database is much more extensive. And even if data is really where is the value, you don’t want to spend too much time on acquiring a database, even at the very beginning. And then you can learn incrementally from time to time, but you need the first database, and you don’t want to spend too much on this.

Additionally, we can say that another challenge that people may face in the digital olfaction industry comes up – it’s funny, there is a flattering perception coming from the first impressive success of deep learning in vision or speech recognition… And for sure, tools working very well for vision. The odor sensor will take benefit from it, for sure.

But the challenge is from another type. I believe when it comes to odor sensors, first, even if Aryballe technology can offer an odor - that’s been said during the presentation - within a few seconds and few minutes sometimes, acquiring odors is a bit longer than just taking a photo. And after this, no one can label the digital result, if he was not present when the smell was captured.

If you see a cat in a picture, even two years later, you will be able to recognize it and attach the label “cat” to the picture. With odor signature, it’s not the case. The capture signal doesn’t smell anything.

So I believe labeling data is a challenge for high-volume, and these particularities are a bit slowing down the ramp-up towards big data in digital olfaction, for sure. It takes more time.

Yes, also, due to the particular nature of emerging odor sensor, we made at Aryballe a huge effort to have our sensor as much as reproducible and repeatable along with time, while keeping the discrimination and reserving power through the diversity of chemistry in the odors alive, I would say. Then you can learn the models and make them efficient in real life, like automotive application, or flavor fragments… That’s a couple of usual challenges you may have when doing a smart system with AI, and maybe not only with digital olfaction.

Thank you for sharing those. I think a lot of those data labeling quality issues that I think a lot of industries can learn from that… That’s a lot of times where things slow down. And yeah, it’s so awesome to hear about the transformation that’s happening in both of your organizations, and at this event, just hearing what’s going on, and how these technologies are really shifting people’s perceptions of what’s possible is really exciting. So thank you, Yanis and Mary for joining the panel. It’s been great to chat with you.

Thank you.

Thank you very much.