ai.bythebay: Michael Ludden, State of AI: Chatbots, Robots & Virtual Reality, Oh My!
Recording: ai.bythebay: Michael Ludden, State of AI: Chatbots, Robots & Virtual Reality, Oh My!
[Music] I planned this talk as more of a general you know state of AI talk hence the name state of AI with a few specific examples that I'm aware of that I think you might find interesting I've layered in some plugs for for Watson just because I know this is related and a lot of people will be interested in hearing about Watson but essentially yeah I'm gonna go over a general like sense of the state of AI and I usually like to speak in a pretty collaborative way so if anybody has any burning questions feel free to just raise your hand or shout something out if there's additional information you have to add to anything I'm saying or questions or clarifying questions you have there's a lot to talk about I don't have I've intentionally kind of like ripped it up in a bucket so that we can burn through it a lot of this stuff many of you may know some of it maybe some of you don't so I guess with that said but I'm further ado I'll just start it up oh by the way this is my information in case you want to follow me on Twitter or look me up on LinkedIn or II have a question you want to email me about I run a team called Watson developer labs which is a new team we're staffing up for and more on that in just a little bit I'll give a shameless plug don't worry ok so just the agenda what I want to talk about so I like the level set my dad grew up reading the English dictionary the Oxford English Dictionary so I'm always careful to define the terms before I use them so that you know what I'm talking about and I know what I'm talking about and and then after we do a little terminology definition that'll lead us right into you know maybe the an overview of the different approaches to artificial intelligence that I've seen of late this is this is just me doing you know industry-wide not just competitive analysis but analysis of of the emerging artificial intelligence business community etc and and so take that for what you will and then I'll give you a few examples mostly from the Watson world because that's what I'm aware of in terms of chatbots robots and virtual reality then the shameless plug and we'll take copious amounts of Q&A so if you did want to just write down your questions I'll be sure to answer more than two very auntie okay so terminology first of all who knows who's ever heard the term machine learning I know it says it'll softball here so machine learning is is in my opinion and there will be a lot of my opinion in this so take that for what it's worth what we mean when we say AI right now as far as I'm aware there's not much out there that is unrelated to machine learning nor is have we reached so sorry machine learning is what I would refer to as narrow AI has anybody heard the term narrow AI okay what is it then you wouldn't you nodded like yeah I know like I don't know that's okay no worries noise he raised his hand behind you too what you got that's a good yeah that makes a lot of sense so I agreed narrow AI as he said is a artificial intelligence slashed machine learning algorithms which is what they're made up of that are good at tackling specific problems so for example you can imagine at IBM you know we like to talk about deep blue and jeopardy right well those things were purpose-built artificial intelligences that could really only do jeopardy and really only do chess right it couldn't give you relationship advice or have a child or decide to be interested in art right that's general AI when we're talking about in real intelligence some people say consciousness but there are certain requirements right like it has to be able to do more than one specific thing that it was prescriptively trained for we're talking about emergent behavior that of course conforms to certain guidelines and rules and there's a lot of talk going on about those obviously you guys have probably heard of Elon Musk's stuff lately the universe di that he launched which is essentially an interface for any machine learning algorithm that you can take and you can have it play a number of games or do a number of things in the same way that a human might so that it would learn to manipulate and use the internet the same way we do and manipulate and use different control schemes the same way you do so basically it's like a layer in between what the AI sees and what it does where it has to just like us send keystrokes and understand from a pixel arrangement perspective what's on a page and you can take machine learning algorithms via that have been trained via deep learning techniques or unsupervised learning techniques that you can then train up and it's really cool to see it happen where you can go from hey I don't know what this is - okay I'm really great at playing pacman or whatever it is right it's cool to see that concept which took an entire giant corporation like IBM and more recently Google with tensorflow and what they did with alphago which is a really complex game a very long time an entire divisions to accomplish now we're starting to see it be a little bit more democratized which I'm personally very excited about so that's narrow AI generally I think we've kind of defined by by by defining narrow AI deep learning is a technique by which you train machine learning algorithms to figure out certain patterns right so the easiest low-hanging fruit example that you hear at conferences like this and others is around visual recognition right so the easiest thing to get a machine learning algorithm to do is to do some basic pattern recognition you can teach it very easily for example to recognize what is and isn't a specific picture that you give it at its most basic level importantly for deep learning or other machine learning training techniques and obviously mostly leveraging neural networks that is something that you only have to do when you're training your own machine learning algorithm to do what you want it to do you don't have to do that when you use a cloud-based platform like Watson and I'll get into that in a minute cognitive is a word that at IBM and other places actually other companies are starting to adopt this it refers to generally smart systems smart applications smart ways of doing different job roles it's a pretty broad term it's meant to be sort of an industry standard term that's a bit of a blanket term that you could use to mean machine learning algorithms or any kind of AI right that's trained in any way so that's cognitive and then machine learning as a service artificial intelligence as a service well that's what our platform and some of our competitors are that basically means just like you would with what's another example like imagine if you're making an application you want to use Google Maps well you're not going to rebuild an entire mapping system you're basically going to use an eight provided by Google or whoever you want to choose right and that is a something as a service in this case mapping and it basically gets you a shortcut to the value you want so in our case we have 30 services you call them via REST API they're all available online and if you Google Watson developer cloud you can see what they are and you can basically do things like advanced visual recognition create an entire chat bot do data analytics stuff like that much faster than if you were to have to train your own machine learning algorithm look Li now that doesn't mean that it's the most relevant or applicable in every single situation that you'll come across there are certain instances in which using a locally based personally trained purpose-built machine learning algorithm is actually better than using a cloud-based platform either you don't have an internet connection maybe you're going to space or maybe you just wanted to do one very basic thing and you don't want to be beholden to other companies for what you do there are use cases in which building your own rolling your own makes sense and in fact that's becoming even easier to do than it ever has been before and that actually kind of segues into I'll skip to the sort of other approaches that people use before I go into detail and Watson so cloud-based platforms include us Microsoft's cognitive suite of services and some of our other competitors was I clear on what cloud-based is does everybody understand the difference like basically you don't only algorithm you call the algorithm that we've been training right does that make sense okay and so we're not alone on that locally based algorithms obviously you guys have heard of tensorflow right you guys heard of tensor flow so there's a bit of a misunderstanding about tensor flow that I want to clear up here because again along the lines of terminology this is such an emerging industry it bears repeating right so tensor flow is not the same as Google offering offering for free and as open source the neural networks that they use to train their machine learning algorithms it's not its helper libraries that will help you build neural networks and train your own does that make sense so it's not quite the secret sauce but they did a brilliant job in marketing it as such and and and it's very popular now and it's one of the most I over T it's the most popular at this point unless somebody wants to dispute me open-source helped rely very framework for training your own locally based machine learning algorithm along with that companies like Nvidia have decided and have seen a market opportunity and and claimed it which is great too that they saw that people were using the computational power of the GPU to accelerate GPU training and so they went ahead and made an SDK that will let you do it faster so basically you can do more a be testing more unsupervised learning more iterations faster using the computational power of an Nvidia chipset using their purpose-built that's DK named CUDA with something like ten strip alone they also have it for other ones as well so and then of course there's the open source approach there are really great marketplaces like algorithmic and also companies full companies that are based around building open source algorithms that anybody can use and building a solutions layer around that as well so these are the different models that companies are you know experimenting with now I think there's been enough traction with all of the different approaches that I don't think any of them are going anywhere anytime soon I think we're still in kind of the Wild West of what is quote-unquote the right approach and I think you're gonna see a lot more clarity around that in the next few years and so in terms okay before I go to that so then I guess at this point I'll tell you a little bit about the Watson platform how's that sound Dada good great sounds great according to one person that's enough we have Korra okay so this is the Watson developer cloud suite of services I've put up this rather boring page because it gives you a list of mostly the the buckets that we divided into so you've got language speech vision and data insights and I'm not going to go so much into the weeds with each one but I'll talk through a high-level and you can go to the website yourself and there's demos for every single product that you can actually play around with it will show you the features but also there's a fork on github link that you can actually fork it and do it yourself if you're a developer and you want to test out how it might work in your scenario or alter a few different things kick the tires basically right and then in addition to that of course there's a bunch of documentation you can see details about the pricing but essentially just like every other API platform cloud-based platform we offer this with a dear to all the services so anybody can begin prototyping for free all you do is you sign up for what we call a bluemix account bluemix is our public cloud you can think of it as IBM's version of AWS or Heroku or Azzurro things like that so basically you sign up for that it's free there's a 30-day free mostly unlimited trial after that you still get the free tier of the Watson services you'll also get to check out all the other things that bluemix has on offer some of them are really cool I recommend checking out open wisk and checking out node-red and a few other options there but essentially for our purposes today in order to get started prototyping with this all you would do sign up for a free bluemix account get your API credentials and go some low-hanging fruit is so I want to call your attention to watson conversation watson conversation is our bot training interface it has a UI which is very which is along the lines of things like way eye or converse a eye the Microsoft bot framework Louis things like that and that basically lets you create a dialogue flow with a lot of error correction built in on the cloud side all without you having to do too much work it's really interesting if you get a chance I check it out if you pair that with speech the text in the speech section and or text to speech for for response output then you have a fully functioning chat bot that you can talk to right so conversational experience on a mobile device chat but you can talk to and there's a lot more use cases I think that have been exploited so far in that area we'll talk about that in a sec so then moving on to vision I mean by the way in language there are a lot of other things that you should check out that are interesting natural language class and not classifier personality insights alchemy language and these all give you extra metadata and information about what somebody means how they're saying it why they're saying it what are the important themes and things like that or and even are they angry are they happy etcetera tone analyzer so visual recognition is our visual recognition service obviously but one of the key differentiators of this service and actually our platform that I like to talk about is the retrain ability of it so while you don't actually have to train up the machine learning algorithm yourself and in fact you can't what you can do with this service and a lot of them is build your own essentially I guess infrastructure data system on top of what we which you then own we see this is kind of a differentiator because we're not looking to collect people's information or their data or aggregate things about them this is our business models different and so when we talk about hey you can build this on top of it you can own it we mean that so an example that's brought up a lot with visual recognition is one around agriculture wherein you can obviously off any off-the-shelf cloud vision API will get you to recognize an apple right I give it a picture of an apple it'll say oh this is an apple right but what if you know you you are some company that sells apples to supermarkets right and you need to know what are the Red Delicious apples and the you know granny smith apples and what's the difference okay one's red one is one is green maybe it will get you that maybe not however what if you need to know what are the granny smith apples that have like a weevil in them or that have like a little blemish or that are a little bit malformed out of the tolerance for the shaping etc you can train it and build classifiers and classes within the sub classifiers that will get you to a ridiculous amount of granularity and it very quickly the system works by submitting a package of images you don't even need to submit negative images honestly you do in the demo I recommend trying it out and then you can retrain the service to recognize pretty much anything I did an experiment I did a hackathon last year to see if anybody could come up with anything interesting using our visual recognition service and the custom training feature in virtual reality and I was blown away by what people were able to do there was a team that won that built an open-source tool that allowed users in a virtual reality environment to draw a symbol in 3d space submit it to the system and then have the system match that with an image so in this case it was that the tool is for training the service by the way to do it really easily so in this case it was like a key you can go look this up too you just look up svv our visual recognition hackathon but basically you could draw a key submit it to the service and then a 3d key would pop into the game world because the system recognized a key that you drew now in case you're not familiar with visual recognition services you may not be aware that you really can't give it anything that's not real because it won't know what you're talking about if it did it would represent false positives for everybody else so imagine if you want to understand what a key is a real key in real life right so it needs to be able to recognize what that is and biasing it towards recognizing a drawing of a key in flat 2d space would totally mess up the model does that make sense for everybody else so so this capability as far as I understand it is fairly unique among the cloud-based services and it's it's actually a strategy that we're applying to the different areas there are other services that you can retrain you can retrain services with custom corpuses of data in fact in the conversation example if you want to get extremely granular and really really detailed you can import a custom model for a specific industry so that it can automatically air handle for nomenclature which may not be something that it could necessarily generally understand think about medical terminology flora and fauna maybe even financial stuff like when you start to get really in the weeds of specific industries and you want to make the error handling that much more accurate and you're ready to go to production with customers that becomes a very useful feature okay I've already spent so much time on visual recognition have a look kick the tires the demos freely available it's pretty nice there's also other services called or other features within the visual recognition caught like similarity search so you can search for one thing yet similar images of another thing and other things such as that I really recommend you try it out and then when finally we get to data insights so what's not on this list I need to update this screen shot is Watson discovery the Watson discovery is a new service we launched that basically takes variously structured forms of data and provides insights on it's a data analytics tool I recommend you check that out as well we started to do some experiments with for third-party developers or to showcase to them what would happen if you combine a working chat bot or a conversational interface with the insights you can get from a Watson discovery service what would that be like would that be good for you know that would that be like a ready-made package for retail hey natural language query our inventory here and get insights or whatever so there's a lot a lot a lot of blue ocean there I'm really excited to see what developers do both in concert with other stuff that we have and just in general so that's an overview of our services and developers are using this now it's been out for about a year to do all sorts of different things obviously chat bots are really hot people when they think of and they think of data insights so the Watson discovery service is something we're pushing a lot this year and you'll see this continue to grow a general theme and this is going back to the state of AI with all of our services all of our competitors and industry-wide with artificial intelligence is the commoditization of the base technology layer so no longer is it that special now that you can train your own machine learning algorithm to have basic keyword recognition right everybody can do that and we see that going more and more in the direction of the base technologies are not the special sauce but what is is use case based applications like if I come to you and I say hey yeah you want to do you know a customer care chat bot it'll it'll reduce cost with your company by this much and you'll be able to auto route certain things and customer satisfaction will go up and here it's ready to go you can give it to your developers you can alter and edit anything you know that's the sort of thing that begins to have legs going forward and the more you can lean into that the better we've become and we've already undergone that path like Watson conversation of full bot training service used to be a series of four different individual api's some of which have been sunsetted others of which are simply used as a part of this this more complex purpose-built service does that make sense it's a lot it's a lot ok I'm gonna take that silence that absolute silence as a yes and move on okay so emerging use cases yeah I'll talk about one that we built just to kind of showcase how how not fully settled to these are in the best possible way like it's a really exciting time to be alive and in this industry for sure we built something called the IBM speech sandbox and launched it at CES this year in partnership with HTC for the HTC vive which is a virtual reality headset that you can walk around in and touch stuff in 3d space with touch controllers so I had this built to showcase a number of things one how a cloud-based platform might actually be useful to how you might build a chatbot or a conversational experience in virtual reality and three that you could do it with Watson easily we have a Watson Unity SDK and we can allow users to make use of just two services in order to build this experience so basically it is not called cognitive VR I thought I put the latest one but I didn't it's called the IBM speech sandbox and here's a video of what what it is not sure if we have sound if not it's okay yeah it's so this is basically what they're doing I'll talk over it but it's meant to showcase audio interfaces in virtual reality this is the brief tutorial but basically it's a god mode sandbox demo where you can point anywhere and create modify or destroy objects with just your voice there's a couple of things that are interesting about this voice control using a chat bot training system thinking through how conversational interfaces might be applied within new areas in ways that aren't just slavishly like here's a virtual bot here and here's a virtual bot there so in this one it's essentially showcasing what it might be like to use your voice for certain UI features so there you go and and you'll notice she doesn't have to use a wake word like hey Watson or whatever she's just saying things and it's always listening always streaming so she created a large guerrilla and a small guerrilla by using just a natural sentence and the interesting thing that this highlights about Watson conversation which is kind of invisible is is all of the error handling that it does for you so if she had said create a huge gorilla it still would have made a large gorilla and if she had said a tiny gorilla it still would have made a a small gorilla right and if she'd said make instead of creating it's still what if par status create and those aren't things that are hard-coded that's just the system understanding contextually what make or create or huge is right in a sentence so anyone can download this and there's an accompanying how-to guide so if developers are interested in either AR or VR or whatever the use case is really but this was showcasing the the Unity integration you can go and rebuild this yourself and we're working on as part of a new initiative with the team that I'm building making this a freely available bit of sample code on github yeah I know it's just you can tell this was a developer that liked guns I only fought the battles that I could fight right so there are 200 objects in the game it's kind of fun but mainly it's meant to give you a flavor for typically audio interfaces and how they might be done right or wrong in virtual reality this is not saying this is the perfect way that it should be done it's only meant to show developers the art of the possible so that's one use case folks didn't maybe think of and by the way we're seeing really a crazy amount of traction with non-gaming VR use cases we're talking about training therapy education etc so if you wanted to look at that yourself or rebuild it check out IBM dot biz slash Watson underscore VR that's the how2guide showcase how we did it oh yeah a shameless plug I'm hiring so if you're interested for product managers as a part of Watson developer labs which is my team which is focused on producing developer facing solutions that help them use Watson platforms that are I'm sorry IBM platforms that both put IBM's best foot forward and also solve a real problem for developers so that's a lot there and if you're interested take a picture of that and look up that number okay so at this point I guess I'll just walk through a couple of other use cases that you may already know and then I'll take Q&A so obvious use cases for virtual reality I'm not virtual reality artificial intelligence of getting my wires crossed would be you know obviously any sort of conversational interface I don't think we've struck paydirt on that yet I think I think we would all agree most of the chat bots that are out there are essentially useless or they're one one-time-use oh that's cool and in there they're done but I think that you know I'm not gonna make a prediction like it's going to be 2017 but in the next couple of years I really do think that chat bots are gonna start becoming more useful because they're actually going to be starting to do the hard part which is the backend integration so really chat bots only makes sense if they can do something for me which I don't want to do myself and so really low-hanging use-case fruit that's there is around booking corporate travel stuff right like who wants to rifle through all those legacy systems that we and a lot of other companies have and you have to go in and and and cross-check different things well what if you could query a system that would do that for you and not necessarily take action immediately but bring you the results into that same UI as opposed to you having to log into you know single sign-on go to this page and go to that page and rifle through this and that and I think that that opportunity is there for a lot of different industries and a lot of different use based stuff it's just that you know it's always the last mile right that's the hard part how do I get a chatbot that integrates with all the backend systems is that is it easier if a chat bot can just give me the weather or something well yeah maybe but it's less useful so I think we'll start to see more really well thought out coming out of stealth you know opportunities for for chat BOTS to really be useful in the different forms that they are I think you know optimization of a lot of different processes I think if we're just going to take a step back the way that financial transactions are done the way that laws are even implemented and legislated the way that different aspects of our world are monitored and insights are gained is obviously going to change I know that's really broad but you know it's it's it's machine learning algorithms and artificial intelligence are gonna start to affect practically everything we do they already do in ways that we don't even notice I mean a lot of times we you know if you're using Google inbox that's something that is powered by a machine learning algorithm that's going to tell you what's important and what's not you can choose to that makes me a little uneasy I don't like when things are happening invisibly in the background it's already happening with ads with auto-generated sports recap articles those are a lot of times done by machine learning algorithms or just even standard you know Excel sheets that decide what goes in there right like we should always say who won and here's an interesting two statistic so things are gonna start to happen in a more invisible way that will automatically personalize more and more of our lives for us I hope that takes a an enlightened approach as opposed to a an invisible approach that's a little bit something we don't notice until it's already too late but those are other use cases use broad scale so yeah that's pretty much what I have are there any questions no questions Wow nothing thank you Michael so I wonder have a lot of folks in my meetups who basically are startup developers they design this and they basically need let's say Chuck capability for their app so how should they engage with your program to use this API so to build a product such as you know a chart for customer service support right so the question is about hey a lot of people want just a chatbot right for their app or experience or whatever how would they engage with us to do that okay so the first thing I would say is have a look yourself not necessarily from your developer just yourself with the Watson conversation service there's an overview video there's a demo you can check out since you asked I guess I'll show you the demo or or I'll show you well yeah I mean so work I want this demo to be updated but for now I guess it'll give us one hi it looks like a nice drive today what would you like me to do it looks like the lights are already off I understand you want me to turn on something so you see this you see why context is important right so one of the best practices that should really be implemented for every chat bot is that it should remember what you said last and use that as context but of course you need some sort of cloud storage for that and a database that you've selected it really is not all that difficult but a lot of experiences just don't do that you know and it's very annoying if you're a user because it's something you expect so turn the lights on and here's a little highlight make the lights go dark make the lights go off okay so that's a bit of the error handling in effect there now it's not perfect obviously but it's important to note that this response was not hard-coded in there so you're meant to give a minimum of five example responses for a given intent when you train it using the UI and really the UI is meant for anybody it doesn't have to be developers it's it's essentially a visual dialogue tree that you can string things together anybody can understand it and it's pretty cool you can even test it in the UI so you can test the changes you've made or the assumptions the systems made and you can alter them but what's important is that you know some of the nomenclature is not going to be there and it can automatically parse for that and take action on a user's behalf and you can actually go in as the application developer on the backend and see all of your user interactions and see which ones are false positives or where they were given the wrong one and then go ahead and manually rebuy Asit you don't just have to wait for it to get smarter on its own which is another nice and somewhat unique feature so it also ends up as sort of like a data insights or or a dashboard that gives you information about your users as well an analytics dashboard almost and we're working on new features for that too does that explain a little bit about how we think you might get started building a chat pod okay and that's the back end then of course you can deploy a front end to whatever you want whether it's a mobile application or a web UI interface that's up to you yeah so you're mentioning that you know you guys weren't collecting data using this why not yeah so let me click let me let me clarify that so we're not collecting data about the sub classifications that you determine are important to your business for example with the the visual recognition Agriculture example right so we're not going to give to your Apple selling competitor the same subclasses that you you created those are wholly owned by you and not brought back in existing now are we always retraining our machine learning algorithms and making them smarter absolutely we are yeah but that's that's the base machine learning algorithm that you can take off the shelf and use as is and so can everybody else and I don't think there's a reasonable expectation that we would not make that smarter because that drives a value for you and of course we need to have as much user input as possible in addition to it just sort of scraping the internet all the time for images and whatnot does that answer your question all right and I totally agree the commoditization of this base layer of technology is sort of table stakes these days what are some of the quote unquote special sauce combinations of api's or capabilities that you've seen that you think are really cutting edge and potentially you know like that next step for words for AI oh wow okay yeah that's really interesting good question so I mean this is boring but the speech-to-text and conversation makes you get a chat experience you can talk to and it can understand you that's kind of like magic because all of a sudden you have you've reached uncanny valley and you can start you know talking to things I think that's especially true with virtual reality at the presence you're there it's it's it's a very powerful thing if I if I take my my IBM cap off a little bit I'm really interested in what happens with self-driving I think that self you know in videos essentially giving away all of the stuff that anybody in this room would need to build to roll their own which is crazy when you think about it but I mean if I can go ahead and go train a machine learning algorithm which can steer my car you know from a software perspective I think that's very interesting that's not necessarily a combination of things but you could imagine that that there would be different artificial intelligence agents at work and a more complex system that would work together I think that's a very interesting thing as well I made a joke last year when I was doing a training class about a tinder for algorithms like where they can just kind of meet and have algorithmic babies together and it's silly but it's kind of it's kind of you know it's gonna happen and say in one form or another it may not be called tinder for algorithms but a grinder I don't know but yeah let me let me think on that there are a number of really interesting examples of usage of different machine learning algorithms together or in concert that do really interesting things but I'm blanking at the moment so let me think on that anything else or I have no time to think yeah you [Applause]