What Watson does | Alyssa Simpson

And I'm gonna shift a little bit to kind of what does this all mean and what do we do at Watson every day. So who in this room knows how to call an API? All right, maybe about a third of you. All right, does everyone know what an API is? Mm-hm. All right, cool. All right, so this is the Watson Developer Cloud. It's a series of APIs. All of these APIs do different things and it depends on what data you have available and what insight or action you want out of it. So I oversee a bunch of these, but they're broken down in kind of things that have to do with language, right, text. Things that have to do with auditory information, right, speech, me talking right now. Vision, so pictures, videos, right, any kind of image data. And then data is a little bit of a catch-all here. We have a news service and trade-off analytics around making decisions. And these are basically like Lego blocks where you can create applications and hook them together in cool ways to find out cool stuff. So the idea here is that any one of these APIs alone isn't particularly interesting, but when you stack them together like Legos, they do really cool stuff. So I'm gonna show you a couple examples. Any questions so far? Yeah? For your speech-to-text API and your text-to-speech, what languages do you offer? I don't know off the top of my head, but if you click on that right there, it'll— I think we offer 7 or so. So Spanish, and a couple different dialects in Spanish, possibly Arabic, possibly a couple others. But yeah. Each API is at a different level too. And so we have a full breakdown, and more coming this Q4. There are a lot, especially conversations in Chinese and other things. So that's happening at a different level. I'm pretty sure Japanese and Korean are Q4 for speech-to-text. It's really awesome. It's not robotic at all. But my question is, for the recognize language, recognize voice, I mean, for Chinese people, we speak English with a strong accent. Yeah, sometimes the API struggles with accents. Yeah, that's a good example of something that we just haven't done well yet. We may say we support English, but if you speak English with a strong accent, the API might not perform from an accuracy perspective in the same way that you would expect with a native English speaker. So if you guys have more data, or like a Chinese-English— Exactly. Yeah, so we're in the process right now, and it's my job and the team's job to improve all these APIs to continue to cover more use cases, more features, more functionality, more languages. And each one is a little different in terms of kind of where we are in that journey and that process. But yeah, perfect example. All right, cool. So I'm gonna go through some examples here. Anyone ever used Autodesk? Yes. Anyone ever called their call center frustrated? No? All right, well good. That means they're doing a great job and they're listening to their customers, listening to your feedback. So they're an example of a client that's using Watson to lower their support costs. By driving away things that are easily automated, kind of low-level, easy-to-answer questions, using this cognitive technology to figure out what are clients likely calling about, can that be addressed easily, and saying, hey, let's direct you to the right answer here so you never actually have to talk to someone to waste their time. Every time you call a call center, on average, it costs around $9 or $10 per call to that company. And so companies are really excited about reducing call volumes. I'm so happy when I don't have to talk to someone and I can just solve my problem immediately. So they have lower resolution times and they're increasing their capacity. Similar example with a large bank in Canada improving accuracy around, so to your point around accuracy, they had a model, and this is getting cut off here, but it's 97% accuracy that they achieved in their particular benchmark, in their particular model around how do you answer this question correctly. Faster results and delivering customers the answers that they had. Anyone ever done any legal research? No one. All right, a little bit, maybe? Okay, so I have a bunch of friends that went to law school, and after they graduate from law school, their first jobs are usually a low level in an attorney's office doing legal research, Huge pain, right? And so for any case that comes before that firm, they need to look at any kind of prior cases that are similar in any way and figure out what was the ruling in that case and was it similar and does the law apply here? Huge pain, really time-consuming, and just out of college, each, or just out of law school, each person is paid like minimum about $120 grand, and so it's very expensive research to get done. And so if you can help the, and it also sucks. When you first graduate from college and you're doing this research, People are so unhappy doing it. And so if you can speed up that process and make them easily able to find all this information that they're looking for that pertains to this one case and compares it to all the other similar cases that have come before, you can really improve the value there. So this is a company actually that started up trying to address this problem using Watson natural language processing. Anyone have a Moleskine notebook? Moleskine. Moleskine, yeah. They're like these fancy kind of nice notebooks. So they were running a campaign to increase user engagement. So when you go, they're running a campaign called #MyMoleskine, and they were encouraging people to upload pictures of them using their Moleskines, doing all different kinds of things. And what they were trying to figure out is how to drive engagement and how to increase this, but they wanted to know what are people posting pictures about? So typically, right, if you take an image and you try to put it in an Excel document and organize it, it's kind of hard, right? Images don't lend themselves nicely to tables and columns. And so what we sort of worked with them on is they trained up visual classifiers to distinguish the types of visual data that they were interested in looking. They wanted to categorize it. Is this someone who's journaling, right? Is this someone who's doing landscape type of drawings, right? Or are they taking a picture of their notebook? Or are they drawing a portrait? Or are they just uploading other random random stuff that has nothing to do with Moleskine. And so they trained up classifiers in these categories and then started connecting users with other similar users who were uploading similar photos, and they really saw their engagement around this campaign go up. So kinda cool. And to your point around accuracy, accuracy is really in the eye of the beholder around is this a landscape photo? Well, Watson's giving a confidence score of about 0.91. I don't know. Do we all agree that's landscape? On the training data that they submitted, that's what Watson is going with around that's a landscape. So it's really all to do with the quality of training data that is predetermined as this is landscapes. All right. So you guys are in a non-traditional type of school environment. There's another one that made some news recently in the Wall Street Journal for Georgia Tech. So professor was hosting a computer science class, large class, a lot of people online, doing cool stuff, had about 10 different TAs over the course of the semester answering computer science questions around whatever it was that they were teaching, and then at the end of the semester, he actually revealed that one of his TAs, Jane Watson, who had been answering emails and forum responses and everything else, was actually not a human, but actually a chatbot, essentially, built on top of Watson. So kind of a cool use case. Freaked everyone out being like, oh my God, I've been getting emails in the middle of the night from Jane Watson answering my questions, and his point around accuracy, he trained it based on years of past question and answers responses from students and TAs to achieve it to his benchmarks, about a 97% level of accuracy around the answers. What was also cool in this case, and I'll get to your question, was that think about a TA trying to answer 500 to 10,000 students' questions, right? 2 o'clock in the morning, assignment's due early, you're up late, you're annoyed, someone's asking you a stupid question where there's already a response. You may have kind of a quippy, annoyed tone, right? Or your personality, your response might be a little frustrated with that student. And so another sort of added benefit was that they trained Jane Watson to always give complete answers and complete sentences and have a tone of voice that was nice and pleasant and encouraging of these students, where compared to some of the actual human TAs answering questions, they were not quite as nice and forgiving and understanding. So in addition to being correct, it was actually a more pleasant experience, and what was crazy about this whole thing was that no one had any idea until the professor opened it up at the end of the semester, including IBM. We had no idea. This was just a guy doing it by himself. So yeah, what was your question? So my question was, since Watson, well, since human hair is always gonna be something that's around, and Watson relies on the input and training of humans, do you think Watson will ever achieve close to 100% accuracy or even like 99%? So again, accuracy is not one thing, and Watson is not one thing. Your classifier that you're training around answering questions in computer science here is completely different from the classifier Moleskine is training around landscapes. versus something else. Those 2 systems are not the same. They're dependent on those separate sets of data. So achieving 100% accuracy is absolutely attainable for a particular use case, for a particular industry, given enough training data. However, those humans have to agree and give clean data, right? So it's always dependent on the data. Okay, nice. Yeah, it all always comes back to the data. When there's bad data, you get bad results, and everyone's like, why is Watson wrong? It's like, well, you gave it a bunch of crap, So of course, it learned whatever you gave it. Isn't it Jill Watson? Jill, wow, I made a mistake, thank you. Yes, I will fix that. Jill Watson, not Jane Watson. Mr. Boyle was data aesthetics. I'm sorry? Mr. Boyle was data aesthetics. Oh, cool, all right, great. I will fix that, thanks. Another one, United Way for graded For Greater Austin, anyone familiar with United Way? They're a nonprofit, huge nonprofit in the United States. They have chapters in basically every city. They do a lot of great work around social services for disabled, needy, low-income groups of people, and extending the reach of their services to people who, serving a much broader audience than they would have been able to in an anonymous way where they didn't expose the personal identifiable details of the people that they were treating. So pretty cool to be able to target new people. Another one here is since we're in California, anyone know we have drought? Bit of a problem. The state's mandated that we have to reduce our water consumption. So this is a really cool example, one that I've worked closely with. OmniEarth is a company that is taking publicly available satellite data, free data that you guys all have access to from USGS, and then they're combining it with mapping data, and they're taking a parcel-by-parcel aerial view. So they're taking these big satellite images and they're breaking it down into parcels. Which land is what, and who does it belong to? And then they're applying visual recognition to understand is there a pool in your backyard or not, 'cause a pool is a good indicator that you're using a lot of water. And the other thing they're looking for is solar panels, because solar panels on your roof typically means that you're a little bit more of a water-conscious person. You may have low-flow toilets installed or other types of faucet and best practices, right? Because they have, these governments, right, local governments here in California have a certain amount of marketing dollars that they need to reach out to you at home and say, hey, use less water, right, please. But they only have a certain amount of money and can only reach out to a certain number of people, a certain amount of time. They're trying to figure out who is using the most water, because it's a tax. It turns out, shockingly, water is not metered particularly well. So it's metered often at a block level. It's not down to your specific home and how much water you're using. And so it's hard to tell who's using too much water. And so they use visual recognition to figure out who is likely users of too much water that they can go and speak to, and combine that with their actual water budgets and water usage data. So they built an awesome company around that, and What's cool about it, again I'll emphasize, is that a lot of this is free or really, really low cost. They're just applying really innovative ways of using data to add value. 'Cause governments in California, shockingly, are not the world's most sophisticated users of data. Anyway, so really cool story with them. Another one is The Hunt. Anyone ever seen someone wearing something cool, like, hey, I love your shirt, let me take a picture of it. Wanna find something similar? Yeah? So The Hunt is a— what? Yeah, good idea, right? So they built an app, a company. They were actually acquired by Pinterest about a year ago in December around this technology. They were using some stuff put out by us to do this sort of similar image search approach. So they created a database of images. So about 5 million images, and then they built an app where you could take a picture, right? Cool. And then they searched through that database of 5 million images to come up with things that were most similar and then direct you to the website of where you could go buy that shirt, right, or a similar shirt. So they achieved about, I think, a 40% hit rate. It returned a match within about 2 seconds, and with their app and their mobile app, the retention and the usage really went up after they implemented this feature. So cool stuff there. Thanks, Jen.