There are few things in the world I enjoy more than getting a bunch of smart people in a room to talk about the future. Nothing does that better than a startup pitch competition. This past week, I had the unique pleasure of attending a pitch competition at MobilityX in Austin, TX, to hear eight companies pitch ideas on using Artificial Intelligence (AI) and Immersive Media in the mobility space.
Last March, Capital Factory and Daimler-owned moovel North America (N.A.) partnered to create MobilityX, a transportation technology accelerator aimed at fostering the growth of new mobility-as-a service companies. My role on moovel’s R&D team is largely to research emerging technologies. Combining that with my general love of the sometimes-ridiculous and sometimes-inspiring world of startups, the MobilityX pitch competition was a treat for me.
As I mentioned, the pitch competition was to see how AI and Immersive Media can be used in the mobility space. But these are umbrella terms for a suite of technologies, so let’s first take a moment to get on the same page about what they mean.
Immersive Media refers to digital technology or images that actively engage one’s senses and may create an altered mental state.
AI (Artificial Intelligence) refers to the nebulous realm of computers
mimicking human intelligence, such as cognition, speech, decision-making,
etc. In mobility technology, the four most prevalent applications of AI are
Natural Language Processing (NLP), Machine Learning (ML), Autonomous Driving,
and Image Recognition.
The companies that stood out the most for me were in the Artificial Intelligence space – specifically using Natural Language Processing (NLP) and Machine Learning (ML). Three companies using NLP caught my eye at the pitch competition. But before I reveal who these companies are, it’s worth taking a moment to expand on what NLP means in the context of mobility – and why any startup that successfully embraces it should be commended. Also, in my next blog, I will report on startups with a novel approach on ML in mobility.
Why Natural Language Processing (NLP) is So Tricky
NLP has been around since the theoretical underpinnings of computing were laid down by Alan Turing. The famous Turing test is probably the ultimate use case for NLP implementations. By the 70s, computer scientists were saying over-the-top things like, “translation and communication were solved problems,” and, “a fully functioning NLP AI solution was years away.” As is often the case with technology, these scientists underestimated humanity’s ability to be nuanced and difficult.
To get a feel for the trouble in NLP, let’s take a look at that first use case: something that seems simple, but is fairly complex under the hood. For example, let’s say I’m chatting with a chatbot and I say the phrase, “I’m headed to Milwaukee next week.” There are three relevant pieces of information in that one phrase: “headed to”, “Milwaukee”, and “next week”. In NLP parlance, these are called entities (the process a computer uses here is called Named-Entity Recognition), and they all pertain to different sub-routines for understanding.
First, let’s try to figure out what I mean by Milwaukee. If we were chatting at the bar, you might know that I live in Portland, OR. and that I’m from Chicago, IL. (even if you didn’t know that, my propensity for wearing Bears gear would help you out). Now, in your attempt to figure out which Milwaukee I’m headed to, you might narrow it down to Milwaukie, OR. or Milwaukee, WI. Our intrepid chatbot, armed with the same information, might do the same. Of course, that would require the persistence of per-user historical information, and that’s a big, big software system. Without that, maybe our bot checks spelling, but that assumes I know how to spell the right Milwaukee, which, if I wasn’t writing a blog about it, is very unlikely.
So you think to yourself, ‘ok, maybe we can’t use per-user historical data or spelling. Let’s instead try to disambiguate location based on a probabilistic model based on something like population and/or proximity’. That might work for ‘Milwaukees’, but what about ‘Springfields’ or something less-specific like ‘City Hall’ or the ‘Courthouse’? There aren’t enough interns in the world to hire to make and maintain that type of location-based disambiguator, and it’s unclear how you’d shove that into a UI in the first place. To get an idea of how it might work, I just searched for “city hall” in Google Maps, which took me to the Springfield City Hall at Springfield, OR. Google’s utilizing historical information from my search history along with my location to attempt to present me with the most relevant city hall. This is, of course, a nifty UX, but it’s difficult to raise money by telling people you’re going to recreate Google. And there’s also the fact that Google is wrong. I was actually thinking about the City Hall in Matawan, NJ (bwahahahah!).
A recurring theme in the world of AI is that the real world is very big and complex, and making meaningful intelligence for it at scale is rather difficult. After all, we didn’t even consider the fact that I might have been talking about ‘Milwaukee’, the power tools company, which itself has multiple locations across multiple states and brings its own disambiguation nightmare to the experience. Our range of languages and our penchant for arbitrarily naming and renaming things (thank goodness for %s/Bombay/Mumbai/g – know what I mean?) makes a general purpose NLP Interface a difficult prospect at this time.
My Top Three for NLP
AppVuze, the MobilityX challenge winner, set out to recreate customer support software (think the support portion of a CRM) for mobile. Mobile is an interesting space at the moment as it is on the cusp of proliferation with new devices for the Immersive Media space. That AppVuze has focused on making their offering available not only on iOS and Android, but on the majority of the cross-platform solutions in the space show that they get how to make your offering compelling in the mobile space. While Natural Language Processing (NLP) in customer service isn’t entirely new (you can say, “new card questions”, or “main menu”), being able to assign an app-specific context to that chat is novel. AppVuze is focusing on support for enterprise mobile solutions, which, when combined with things like Google’s refocusing of Google Glass to enterprise solutions, has some definite promise in the coming years.
Upswing’s ballywick is in the educational space, and I have to give them credit for diverting some resources to apply their tech to the mobility space. Upswing combines student support – tutoring, mentoring, and advising – with data science. For the competition, Upswing presented us a Facebook Messenger bot based on a simple decision-tree-based UI. While everyone’s puzzled looks regarding Facebook Messenger bots made me a little doubtful regarding the prevalence of its platform, decision-tree-based UIs have been a hallmark of modern gaming and extending them into other spaces seems like a no-brainer. Education seems like a natural fit for multiple-choice domain-specific conversation UIs, and I’m excited to see how Upswing can grow in that space.
Blink.AI is taking more of a platform approach to the NLP space and endeavoring to provide an abstraction for the creation of context-aware domain-specific conversational UIs. While earlier in the product lifecycle than their NLP compatriots, they definitely know what they’re talking about in the space. That depth of knowledge and understanding of how it can apply to the mobility space was sufficient for second place at the challenge. After all, taking something complex and making it simple enough to enable others to build cool stuff is How To Make Money 101, and I wouldn’t be surprised if Blink.AI ended up powering some intrepid chatbot in the not so distant future.
In my next blog, I will look at the startups from the MobilityX competition that are wrestling with Machine Learning that are on my radar – and should be on yours, as well.