The Future of Watson, AI, and Machine Learning - Armen Pischdotchian24:15 with Nick Pettit and Armen Pischdotchian
IBM Tech Mentor Armen Pischdotchian talks with Nick Pettit about the future of the IBM Watson platform, artificial intelligence, and machine learning.
Hi everyone, I'm Nick Pettit and I am here with Armen Pischdotchian, 0:00 and he is an Academic Tech Mentor at IBM Watson. 0:05 How are you doing today? 0:08 >> Nick what a pleasure, thank you, doing great. 0:10 >> Cool, well first I just want to ask you what is Watson? 0:12 >> Great question to start off with. 0:17 Let me bring us back to 2011. 0:20 That's when we had the Jeopardy game, and interesting story behind that. 0:24 Sony Corporation, that happens to own the Jeopardy game, 0:30 the American quiz show, knew about Watson. 0:34 They approached us and they said so, you say that you have a system that can 0:37 actually beat these two Jeopardy champions, Brad Rutter, Ken Jennings? 0:41 We said yes, we do, we have a high confidence. 0:46 It's never a yes or no, it's probably. 0:51 So we have a high confidence that we can beat these two champions. 0:53 They said fine, and we set up a stage at out conference room, 0:56 and the rest is history. 1:01 >> Well, so you did Jeopardy, and then what happened from there? 1:03 >> And so Watson eventually was born, it became practisized. 1:10 It became a product. 1:16 >> Right. 1:17 >> And I'm always reminded of our CEO Ginny Rometty, 1:18 that she has termed the coin augmented intelligence. 1:24 The idea in here is that Watson is helping people and 1:31 machines work together to create knowledge from data that enhances human expertise. 1:37 Watson isn't here to supplant, and remove jobs. 1:45 It's here merely as a second opinion to the doctor. 1:50 In fact, I'm reminded of an example. 1:53 A personal example that I witnessed. 1:57 On the way here from the airport we were, the taxi that I was in, 1:59 was cut off by this giant truck in front of us, and 2:03 we could see the reason was because the truck was trying to avoid another 2:07 car that clearly you could tell the driver was texting and driving. 2:12 >> Right. 2:16 >> And so you had to swerve if front of our lane. 2:17 >> Right. 2:19 >> The taxi driver said to me Armen, you know my brother is a truck driver, and 2:19 they're frankly, truck drivers are some of the best drivers, period. 2:23 They have a preference to drive at night, but they can't. 2:28 They're mandated to sleep. 2:33 They have to log that they have slept. 2:36 >> Right. 2:38 >> So this puts them in traffice during the rush hours. 2:39 So I'm thinking to myself how great would it be if this autonomous vessel did 2:42 the driving on its own, across country, allowed the truck driver to sleep, 2:48 when he woke up, as it got closer to the destination, 2:52 then he would open up the laptop and do inventory. 2:55 Nobody's job has gone away, but the mundane, 2:58 the arduous has been taken away, so now we can focus on what is meaningful. 3:01 >> Right, so you call this augmented intelligence. 3:07 How do you see this being used in the real world? 3:10 >> I think one place that it's becoming prevalent is in the field of data science. 3:13 Let me take a step back in here. 3:20 If you looked at data, 85% of it is unstructured. 3:23 These are tweets, emails, MRIs, 3:30 CAT scans, audio conversations, videos. 3:33 That is where I want to gain my insight, my knowledge. 3:40 Not so much from structured data that happens to happily reside in 3:46 a table that I can call it with a SQL call from a table or a column. 3:51 Now, structured data, transactional data also has its importance. 3:58 I think where these two come together is in the field of data science. 4:03 >> So what it sounds like is that Watson is able to elevate the level of questions 4:07 that you can ask of this data. 4:12 >> Yes, let me add to that. 4:13 It is accurate, it is accurate. 4:16 In fact, some of the services that we will have in this course, the conversation and 4:18 discovery service, lend themselves to that insight extraction. 4:23 I always like to think of the conversation service, the chatbot that we would build, 4:27 about imagine all the things that you know that you don't know. 4:33 Well, wonderful. 4:39 We can then use a conversation service to find, 4:40 at least you know what it is that you don't know. 4:43 Now think about all things that you don't even know you don't know. 4:45 That's insight. 4:50 I think that is where the discoveries surface, 4:52 especially if it's connected to the conversation service. 4:55 We will explore that in the course as we move forward. 4:58 That's where insights are extracted where understanding and 5:02 knowledge is extracted from information >> So 5:09 tell me about the future, you're talking about the present and 5:13 what we can do right now, but where is this all going? 5:17 >> Currently there are three major calibers of AI 5:21 >> There is Artificial Narrow 5:26 Intelligence, and that's where we are today. 5:28 The most fantastic technologies at work, 5:33 whether it is self-driving cars, or the work done by Amazon, 5:37 Twitter, Facebook, IBM, Google, we’re at Artificial Narrow Intelligence. 5:43 To have a better understanding of where we are now let me take a moment and 5:49 talk about where we may find ourselves tomorrow. 5:53 Artificial General Intelligence, AGI, this is where the machine is equal, 5:56 and perhaps, millions of times further ahead than a human. 6:03 Case in point, an example, Carnegie Mellon University, 6:08 CMU, Robotics Department. 6:14 I had the pleasure of being invited there for a lecture, and 6:17 when you walk into see a muse robotics division, you're greeted by a robot. 6:22 This is not a humanoid robot. 6:26 >> Right. 6:28 >> Cameras, laptop, and a pod, and track wheels. 6:30 I had registered, it knew me, hello Armen, welcome. 6:34 What a refreshing greeting that is, Dan, can I help you? 6:37 But since my bio was available, so it had face recognition, 6:42 it clearly knew it was, and it takes a look at the appointment that I may have, 6:47 and I was told later that if I have arrived early enough, 6:52 It would suggest if I wanted to go to the cafeteria for refreshments. 6:55 In this case I happen to be just in time to get to my meeting, and so 7:01 I asked the robot, can you take me to John's office? 7:06 The robot said, sure, follow me. 7:11 >> [LAUGH] >> So I followed the robot. 7:12 It said on the screen in 20 feet, it will turn left, and 30 feet, 7:14 it will turn right. 7:18 Fine, I did as it did. 7:20 As it went by the elevators, one of those software engineer types, 7:22 neural network software engineer types, kicked their chair in front of the robot. 7:26 It stopped. 7:31 It sees the chair, it stopped, laser, everything, 7:32 nothing that's using technology, that part. 7:37 I stop as well, looked at the software engineer with somewhat of a puzzled face, 7:40 and it went around the chair, and then I went around the chair, and 7:45 finally we got to John's office. 7:48 John asked the robot two questions. 7:51 Two very difficult questions for any machine. 7:56 What happened at the elevator? 8:01 Why were you late? 8:05 >> [LAUGH] >> The system is gaining an awareness 8:06 of its surrounding, of its environment. 8:11 It's gaining experience. 8:14 We're doing similar work at our Almaden Research Facilities in California, 8:16 where we call it embodied cognition. 8:22 IBM Fellow Grady Booch, and his colleagues are working on exactly that. 8:25 Systems that gain experience, and this isn't memory. 8:30 That brings me to Artificial Superintelligence. 8:34 I think one example would suffice to say, consider this. 8:38 If I ask Watson today, Watson tell me about neural networks. 8:43 Watson would say, I'm sorry Armen, I don't know about neural networks. 8:48 Perhaps you can teach me. 8:53 Fine, we'll ingest a PDF or crawl the web, and he will learn. 8:55 As opposed to the day after tomorrow, if I ask Watson, 9:00 Watson tell me about artificial networks, and 9:04 Watson would say something more like this, I'm sorry, Armand, 9:08 I know that I don't know about artificial neural networks. 9:14 Perhaps you can teach me. 9:19 That's self-awareness. 9:21 Suppose you might ask when is this going to come to bear? 9:25 Ray Kurzweil, at Singularity University, is currently at Google, 9:29 a futurist, predicts that AGI would be around 2040, 9:35 at our doorstep, and 9:39 he thinks ASI would be somewhere around 2060, for systems to have that awareness. 9:43 >> So data science is rapidly emerging as one of the most 9:49 important career paths of the future. 9:52 What does a data scientist need to know? 9:55 >> Perfect question, absolutely, absolutely. 10:00 I'm reminded that of Professor Mortissa Bidar from 10:03 Ryderson College in Canada via University of Toronto. 10:08 I think he summarized it perfectly, he said that. 10:14 Curiosity is the beginning stepping stone for a data scientist. 10:19 Is two bedrooms more valuable, or is having more land more valuable? 10:26 Just that sense of curiosity to question everything in their surroundings. 10:32 Knowledge of various machine learning methods helps. 10:40 Although he also is quick to point out that data science is a team sport. 10:47 One doesn't need to be master of all. 10:53 You can pass the baton from specialty to specialty. 10:56 But it helps for a data scientist to be versed in some 11:00 languages that data scientists use. 11:05 Python, they're especially working with Jupiter Notebooks, 11:09 R, Skylab, depending on various applications that they want to. 11:16 For example, R lends itself very well to heavily statistical work, 11:22 regression and so forth. 11:28 Python is a bit more of a general language and it's an easier to master. 11:30 And it also lends itself well to machine learning models. 11:37 And data scientists would detect the data that they have collected. 11:43 And they would run it through various models to try and 11:48 see what sort of accuracy each model gets. 11:51 They will try things such as support vector machines, they may try random 11:55 forest trees, Bayesian inferences, I call these brute statistics. 12:00 The core logistical regression is the engine that's running. 12:05 If their data happens to be multimodal, a lot of variables, 12:09 they would perhaps try neural networks. 12:16 With images, 12:20 it lends itself that much more readily to convolutional neural networks. 12:21 CNN's, of course, have their own drawback but 12:26 the diminishing gradient problem. 12:29 Imagine when you take a copy of a paper, and then you take a copy of that copy, 12:34 then a copy of that copy, eventually it starts to become not so clear. 12:41 Perhaps by the tenth copy, you can't even read whatever it is. 12:46 That somewhat is akin to the diminishing gradient. 12:52 The first layers of a convolutional neural network 12:56 train much slower than the latter layers. 12:59 The first layers might just look for edges that would say that this, 13:03 my chin is an edge, doesn't know it's a chin, that's an edge. 13:07 A layer later, 13:11 we would found two other edges that always appear in the two sides of this one edge. 13:12 So those are the layers that launch lower, and it perhaps lends to a little 13:16 bit more inaccuracies, but then the other layers take that, inherit it. 13:22 So they might find that perhaps they want to use another method called recurrent, 13:27 or recursive neural networks, recurrent or recursive. 13:32 It actually treats each layer as though that's an input node. 13:37 So it starts over freshly, it has a feedback mechanism. 13:43 They might even want to try without their algorithm's Long Short-Term Memory, LTSM. 13:47 We're working with that quite a bit at IBM to help gain the system an experience. 13:53 And of course, a scientist then they may pass the baton or 13:59 the professor hide their stresses that the visualization, 14:05 the explanation of what it is they have collected. 14:10 What it is that they have built patterns out of. 14:14 After all, machine learning is a fancy way of saying build me patterns. 14:17 And that visualization is what they then need to pass on to the c-level executives. 14:22 Or to their end users to impact the decision that they need to make. 14:28 >> Very cool stuff. 14:34 So you mentioned that we're going from, Artificial 14:35 narrow intelligence to artificial general intelligence and then super intelligence. 14:41 What are the steps along the way between each one of those milestones? 14:47 >> A wonderful question. 14:53 Let's go a little bit further back. 14:54 In the beginning, and I wanted to quote Pedro Domingos, 14:59 in his wonderful book, The Master Algorithm. 15:04 He depicts stages. 15:10 Stage one is calculations. 15:13 That was perhaps in the beginning of the 20th century, let's just think about it. 15:16 Even prior to that, mathematics. 15:22 That lended to stage two, methods and algorithms. 15:25 Think of C++, the object oriented program. 15:29 That lead itself to stage three, analysis regression, tried and tested. 15:33 It led itself then to stage four, supervised learning or optimization. 15:41 Very much a machine learning methodology used nowadays. 15:47 Supervised learning has the benefit of you not needing to have so 15:52 much data, but it is arduous because you have to label that data. 15:57 Oftentimes, it leads to greater accuracies than your other methods of learning. 16:04 A little bit more challenging is unsupervised learning, 16:11 that's problem solving. 16:14 Think of the Amazon recommender app, or Netflix, 16:16 the recommender engines are unsupervised learning. 16:19 I may buy a certain book and 16:23 in the bottom it says others also purchased books as such. 16:25 Now Amazon doesn't know my habits of what books I like or so forth. 16:31 But it has a lot of data of folks who purchased this particular book 16:35 have also purchased these other books. 16:39 Perhaps one of those might be something that I'm looking for. 16:41 In unsupervised learning, data is clustered, but it's not labeled. 16:45 One might go and label a few and you could then call that semi-supervised learning, 16:51 and that itself is another approach. 16:57 The one that really gets me is stage six, there are seven stages. 17:01 Stage six, We called that 17:06 unsupervised asking, rhetorical learning. 17:10 What do I mean by that? 17:16 Let me give you an example, consider this. 17:18 Let's say, I get on my phone and 17:20 I say, Watson, my muscles are hurt. 17:24 My joints are sore, my nose is clogged, what's wrong with me? 17:30 So Watson goes in the back and does it's data analysis. 17:38 And my phone is the browser is geo-sensitive so it knows for 17:44 example that I happen to be in Massachusetts, it does a spacial analysis. 17:49 It does a temporal analysis, it finds out that it is October. 17:56 It looks at NOAA, National Oceanic Atmospheric Association, 18:02 and finds out that it hasn't rained in a while and the pollen count is quite high. 18:07 It comes back and asks me a question, this isn't something you put in a chatbot, 18:14 this is a rhetorical question, do you have fever? 18:19 I say no, I don't. 18:23 Negation is huge, it's a big algorithm. 18:26 No, I don't have a fever. 18:29 It goes back takes another look at the data that is collected. 18:30 Comes back and says, I'm 89% positive that you have allergies, 18:34 4% positive that you have the flu. 18:39 And you can imagine from there, well, what should I take for my allergies? 18:44 You might recommend a certain drug off the shelf. 18:50 I don't take any drugs at all, of any kind. 18:53 Then it might say, why don't you try the local honey? 18:55 [LAUGH] >> [LAUGH] 18:58 >> That leads us to stage seven as you 19:00 were alluding to it, and that's the notioin of unsupervised action, 19:02 doing, we're not there yet. 19:07 That would be artificial super intelligence. 19:09 It's an unbounded action, unbounded it's not stuck between a zero and 19:13 a one in the table that you put in a graph. 19:17 Well, that's when robots make their own decisions, frankly. 19:21 They have a community and they elect a leader. 19:25 [LAUGH] >> [LAUGH] So 19:28 what will life be like when we just have this general artificial 19:30 intelligence everywhere, even a super intelligence? 19:36 >> Great question. 19:41 Perhaps I'll take a smaller bite and perhaps a more 19:44 immediate impact of this, I think it would be as such. 19:49 We've all heard of APIs These APIs 19:54 are really a RESTful method application programming interface for 20:01 various services to interact with one another. 20:06 The notion behind that is a URL, username and password, or an API key. 20:10 That's how you get from this room, where we do conversation, to the next room, 20:15 where discovery is taking place. 20:18 I'm thinking that the next would be not APIs but 20:22 something called natural language interface. 20:26 Bots will be communicating with one another. 20:30 Frankly, they would be talking to one another. 20:33 A service, not an API, but a brokerage bot that happens to be the restaurant 20:36 would be talking with the weather, would be talking with the golf course. 20:41 And advising me that I should leave the golf course soon because the weather's 20:46 about to change and the restaurant's going to get booked really fast. 20:50 I think another approach might be move over apps, here comes bots. 20:54 The notion of us downloading apps on our telephone, I think the days are numbered. 21:00 Pretty soon, I would just look at my phone and I would talk to it. 21:06 I'll say get me a car, I need to get to 21:09 Charlie's Restaurant and tell me what's the special. 21:15 By the way, is that the rain moving in tonight or later? 21:19 I don't need to download anything for this. 21:23 I want the systems in the background to do the brokerage on their own and 21:26 give me the information that I want. 21:31 Blockchain technology, it's here already very much so. 21:34 I'll give you an example of a Blockchain technology 21:38 that I think might work well handful of years from now. 21:41 In your fridge, the milk has gotten down to its lower threshold level, 21:46 so KPI, Key Performance Indicator, has triggered a smart contract. 21:53 That smart contract is well encrypted by Blockchain technology. 22:00 Apple Pay or Amazon Pay automatically does the transaction. 22:05 The drone drops off the milk on the front lawn, 22:12 the robot in the house goes, picks up the milk, and brings it, and 22:15 put it in exact location at the fridge to restock the refrigerator. 22:19 This entire thing really is a possibility of 22:24 Blockchain technologies and what it can do. 22:28 Nano bots, I think may be coming our way sooner than we think. 22:33 It's quite possible that, I think, 22:37 in my lifetime I will have nano bots in my bloodstreams. 22:39 Checking the salt, the sugar, notifying me on my wearable device, 22:42 might be clothing if not a watch, that our men stay off the salt and the sugar. 22:48 >> [LAUGH] >> Perhaps if a certain outlier is 22:54 detected, it might automatically notify my doctor. 22:58 I could get a call from my doctor that says, Armen, 23:02 I've noticed that your ankles are swelling. 23:05 That has something to do with cardiovascular issues. 23:08 I might say, I thought my shoes were kind of tight. 23:10 Why don't you come in tomorrow, all right? 23:14 Of course, the world of embodied cognition that 23:16 we're working with at Amadin research laboratories. 23:20 But the cognition is where systems now have an awareness of their surrounding. 23:23 They gain experience, some of the algorithms we use LSTM, 23:31 Long Short Term Memory, 23:35 go towards helping them in bringing that feedback loop into the system. 23:37 Allowing you to understand that stairs hurt if I go and get fall off passed it. 23:43 And humans do not like it when I do this the experience is not the same as memory. 23:49 And the sense of embodied cognition will give that to AI assistance. 23:57 >> That's a world I'm looking forward to live again. 24:04 [LAUGH] Well, Armen, thank you so much for being here today, we really appreciate it. 24:06 >> My pleasure Nick, thank you so much. 24:12
You need to sign up for Treehouse in order to download course files.Sign up