The Power Of Machine Learning

hello to you Jeff good morning or good afternoon James thanks for having me I think it is officially afternoon this as far as I can tell if this man I lose track of time now your job title has a lot of syllables in it your General Electric digital's your vice president of intelligence systems I had to practice that I don't still don't think I got it right yeah no sweat well let's let's talk about it and what do we mean by that um you know my background I've spent the last 10 years or so building and running two different advanced analytics and machine learning companies one was based around a very popular open-source language called R where we had the great opportunity to sell into a lot of who you would think of as being the who's who of data driven companies companies like Facebook and American Express and Goldman Sachs and spend a lot of time understanding how these early adopters were thinking about embedding advanced analytics and data within their business processes and products after that one was acquired to Microsoft we really wanted to think about I wanted to understand how you solve this problem and a big challenge of taking these techniques putting them from white board into production well the same time marrying that emerging it with the resurgence interest in that field called machine learning and so wise IO was dedicated to that where we basically were focused on building production machine learning capability around specific applications and you're the CEO of wise IO which is then GG acquired wise io exactly exactly into a unit called assist io am i right in the history there that that's correct it was called a smart assist and they actually renamed it into a company called answer IQ which is providing intelligent customer support solutions to big brands like Groupon and Pinterest and people like that so to come back to your question why is my title now intelligent systems right and there's a very short tangent the GE if someone who was the VP of I told you systems before you or the first person helped okay so you're the first person who had that title I am sure actually I guess no pressure get the bar very low there is a story and there is a reason behind that so one of our core beliefs and we can come back and talk a little bit about what we mean by machine learning how it's being used but one of the things that we learned at wise was that there's a very big difference between using machine learning for exploratory purposes or attempting to apply a machine learning algorithm to a problem versus using machine learning in production for large scale systems that are continuously improving over time and the latter is a very hard it's a very challenging problem that very few companies in the world have mastered and so we started using the word intelligent systems to refer to machine learning running in production for missing mission-critical workflows that that's the reason for that right so I want to talk about how machine learning gets used in the real world rather than actually stopped on do slip a little elementary 101 thing and have you literally define machine learning you kind of just did I want to make it really clear my brain is a little slow sure what per say if we're not think about machine learning I think about you know a software system that that evolves and changes in response to data over time yep is that correct how would you how would you act literally define machine learning it is so the the most basic literal definition of machine learning it's a very broad field which is can be considered a subset of the the broader field of artificial intelligence but machine learning specifically refers to the ability for computers to learn without being explicitly programmed and so there's different subsets of machine learning there's what's called supervised learning and unsupervised learning and what those mean in general it's supervised learning is in particular to feel that we focus on it doesn't mean that it's supervised by humans but rather what it means is that it's learning from past examples so if you think of the world of analytics or predictive analytics or statistics as it used to be called what that relies upon relies upon an expert human being to formulate and test the hypothesis so for example in the consumer world you might have the hypothesis that a lower price leads to more sales and what you would do is you would have this expert human being formulate that idea and then test the relationship between price and the amount that was sold what machine learning does and supervised machine learning in particular is it flips that paradigm on its head and what it does is it looks at the historical results and it looks at what were my past sales and it infers the relationships of the variables that drove those sales and so it basically takes that process and it reverse engineers the process of statistical modeling where it becomes really interesting is that now you've removed that human bottom bottleneck in formulating those hypothesis so machine learning becomes very useful and very important in large data and large variable environments where that complexity overwhelms the ability for humans to interpret all of that data and all of that signal right and I guess in some hypothetical future world the machine learning system would not just of all but would literally begin to teach itself and we get into our artificial intelligence in the singularity and all that which is another topic we've been doing any other topic but certainly been some interesting stuff there around deep mind in Google's work around go and alphago and some of the self program in there and in particular that's another field you know leveraging something called reinforced depart let's talk about some real-world examples of machine learning of I know that correct me if I'm wrong she works with Twilio MailChimp Zendesk zenefits what are a couple examples of how they're really using machine learning in the real world yeah sure absolutely so those examples were customers of ours or customers of Weis and we were an independent entity and what we were doing was we were providing intelligence driven customer support solutions so if you look at those companies that you just mentioned what do they do and what do they use they use a customer support platform like Salesforce are like Zendesk and in the case of somebody like a Groupon they get millions of customer inquiries every month and they have several thousand customers for agents around the world who are responsible for opening those inquiries reading them understanding them and then deciding who do I send this to how do we reply to the customer and how do I resolve that and so what we had built is we built the machine learning platform capability to embed within those support applications to both augments and automate the decision-making of those customer support pages and so what we then did is as you mentioned previously we spun that product and solution off we've now taken that same paradigm for intelligent decision-making within existing workflows and we're applying that broadly across the GE industrial businesses isn't one of the problems or one of the challenges in the you know as the system of balls that a human has to evaluate how good a job it's doing so if the system is going to begin to replace those humans that are open to the emails or the queries doesn't as some some level human has to evaluate how that is going yeah absolutely and there is tremendous both art and science to that and one of the areas that are spending a lot of time now is thinking about the concept of user experience design in the context of these intelligent systems because that feedback loop that you describe is so critical to having these learning systems actually learn and improve but if you require the humans to provide explicit feedback in order to make that happen is unlikely to be successful and the humans are unlikely to do that so we believe that there's I'm sorry because it's too slow or but what is it does it make it unsuccessful because it's too slow or what makes it unsuccessful what makes it unsuccessful is the fact that if you take somebody out of their typical workflow and ask them to explicitly provides feedback unlikely to do so and so we believe that you get more feedback and better feedback if you naturally integrate this into the products and solutions that people are already using mm-hmm so say there's a company out there and they've heard about machine learning they understand it they want to get on board what would you say to them like how do they how do they use it how do they leverage it I thought of how to actually get their hands on it yeah so I give really the same two or three recommendations for how to get started and first of all there was a great article that came out and was written and I think HBR our business review a couple months ago and the title of the article was if your company is not good at analytics you're not going to be good at artificial intelligence and the spirit of that was everybody wants to think that machine learning or AI is a silver bullet that's going to solve these very hard problems these big data problems but it's really not and so the steps to get there our first machine learning works well on large scales of data that have been aggregated together that doesn't have structured data but there needs to be a good structure to the way that data is stored and aggregated number two is if you look at the value chain of how one of these production machine learning systems comes into existence there's four things that are needed there is computational power there is algorithms there is this unique data that we referred to and finally there's domain expertise and so the second thing that's needed is a deep understanding of what that data means in the context of relevant business decisions the third thing then that is necessary is an idea of what the good outcomes and the workflows are around that data relative to the domain that is being solved for and if you can start with those three things and start to help have people make and record and document basic decisions on top of that data it's then possible to layer on top machine learning that can start to take over and mimic the human decision-making that occurs on top of it but I'm guessing even with that expertise for many companies they're gonna look out of house for for assistance with that I mean is it is it realistic to think they can go through a large cloud provider and and use one of their machine learning tools I think for so let's come back to that value chain that we just talked about of computational power algorithms data and domain expertise so the use there is for those first two absolutely you can look to Amazon or Microsoft or Google for the computational power in the algorithms however one of the things that we've learned and the industry has really started to learn is that those parts of the value chain are starting to be commoditized and they are a necessary but insufficient component of delivering a value creating solution so yes the recommendation would be to companies to think about leveraging what is available as much as possible but if you assume that that is the case you still need to be able to have access to your own unique data and domain expertise knowing how to apply those tools and the latter is non-trivial and very challenging the upshot where I think the industry is really going is there's a tremendous opportunity for companies to be very thoughtful about their bill buy decisions and how they're consuming machine learning driven solutions and in particular my personal belief is that we're going to see more and more companies be thoughtful about only building at those areas which are a unique core competence to their business and more and more consuming machine machine learning powered solutions at the periphery where they don't have a competitive advantage interesting okay well I mean to put it in context is it correct to say that the machine learning sector is still so nascent that it's a as much of a dream and a hope as it is mostly a reality in business today yeah I would answer it this way which is we are extremely early just were extremely early in extracting real value out of the big data wave that was occurring five years ago where we are is I believe that what we're starting to see is a maturing of the tools what we're seeing at the same time is the very early edge of solutions that are leveraging those tools and what I think we are going to see more and more of over the coming years both on the consumer side as well as on the industrial side is solutions or products that are delivering outcomes to the customers on data that has been aggregated a large a across a large customer base and is then turbocharged by machine learning or AI and we've started to see examples of that in things like the way Buber drives network effects in things like uberpool the what Google does with Google Maps for example and taking how they think about their business model and their data driven business model powered by machine learning and applying that to many many other sectors I think we're gonna see become very powerful how is for example that were a pool example how is machine machine learning driving that yeah absolutely so you could take the the example of by understanding and recognizing the history of all of the trips that an individual user has made where they were when they called so I'll take an example I live in the Mission District of San Francisco and I work in the financial district the ability to look at understand that three days a week I'm likely at 8:00 in the morning to call an uber to go from the mission into the financial district is pretty interested but the ability to not just look at me as an individual but all of those users within San Francisco and now able to ask probabilistic questions to say oh it's likely that I'm likely from A to B somebody else is likely to be nearby what they're able to do is provide me with a shorter trip or less money which entices me to take more trips which gives them more data about where I am likely to be and as more and more of that happens they can then apply these machine learning driven algorithm Edwards on top to optimize those trips and those outcomes and the interesting thing about that is that that could not be done with humans alone you need an intelligent software system to do that unless you had 10,000 people on staff they could just sit there and do that but it would not be practicable so in essence the points to the fact that machine learning provides an exceptionally critical competitive advantage that's right that is going to make a big difference it's exactly right and what to me is so fascinating now having been apart of GE for almost exactly one year we just celebrated our one-year anniversary is that we can start to see and apply those same patterns around data-driven business models and so one of the things that's been very interesting to me is as we were thinking about coming on board people always said well it's going to be really challenging because GE isn't a software company certainly trying to become a software company but what was really fascinating to me as as really understood what this company was about and where we make tremendous amounts of money is in our services businesses so if you take our aviation business unit for example we sell tremendous amount of jet edges and we make a reasonable amount of money on those jet engines but we make a tremendous amount of money helping to manage and service those over their next 20 or 30 years of their lifetime and the way that we did that was very much in the same way that Google or uber are providing data-driven outcomes as a service we were just doing it on an older software stack that now as we start to think about things like predicts our Internet of Things platform things like our intelligent systems platform we can help to modernize and turbocharged those businesses that have been so successful so so profitable for this company well let's look at this sure if we're if we're having this conversation two three years out we're talking about machine learning where is it going and how can a business prepare for where it's gonna be say twenty four thirty six months from now yeah absolutely so the first thing I would say is it's really going to be inevitable that these technologies will start to become more pervasive and they will start to become more embedded in all the products and services that you consume how to become prepared for it is much like we were discussing a few years ago which is first and foremost is to understand the business processes within your organization and the data that drives down because it's not possible to apply these techniques to them without understanding how people are making decisions what the outcomes are and what data drives that so the combination of those two people doing more and more of that as well as starting to see these become easier to use easier to consume I expected to be pervasive across business processing across industries in this sort of three to five year type of timeframe it seems like it's going to be a necessary competitive tool look if you don't if you're not wielding it if you don't have it you're gonna fall behind so it's it's like it's not gonna be optional in one of these years of it I think that's absolutely right I would describe it though as necessary but insufficient and one of the things that I always recommend when I'm talking to to peers or even friends at other companies is that the tools are important but what actually think you know frankly interests me more these days is thinking about the business model around these tools because if not just that having a machine learning algorithms or capability is good but rather it's just like we were talking about with uber or just like we were talking about with Google for example around spam detection and Gmail these systems work better the more data they have to be applied against and if you can figure out a business model that entices people to provide you with more data in return for better results that is a very powerful competitive advantage that companies can take advantage of sounds good Jeff thank you very much for your expertise today I'll send you the link when we're done we can tweet about it thanks Ali well great talk to you thanks for having me

This transcript was generated automatically from the video's captions and may contain errors.