**Click here to view the webinar**
I am the CEO and founder of Infrrd. I am going you walk you through an overall plan of how to bring AI into your enterprise in the new year. Briefly I will be talking about three to four main points.
- Is AI really here this time? – We have been hearing about AI for a very long time. We will take a look at what has changed in the last few years and is it a serious vet this time?
- Open up AI in context of application and software to understand what does it mean for enterprises? – A lot of AI use cases we see and hear about are consumer use cases but there is a lot more work that is going on on the enterprise side. We will talk a little bit about that and see what AI means in the context of enterprises.
- I’ll walk you through some example of how different enterprises using AI today. – Some of them are our customers, some of them are general examples. We will see what is happening on the AI side for enterprises.
- How to go about bringing AI into your Enterprise? – We will talk about what we should do to bring AI into your enterprise, where to start, what are the things to watch out for and how to go about creating your first AI solution.
Before I start, I want to spend a few minutes introducing our company. We are an Enterprise Machine Intelligence company. We work on bringing AI into enterprises across Retail, Financial Services and Read Estate Domains. Our Platform has three fundamental capabilities. Computer Vision, which is our Vision.ai platform, Natural Language Processing and Predictive Modeling. We have canned solutions that we sell to our customers based on these capabilities. We also do customized solution development for a lot of out customers using these fundamental capabilities. A lot of stuff, I am going to talk about today is coming out of our experience working with our customers across these different verticals.
To begin with, I want to talk about AI. If it is really here this time. We have heard about AI since my childhood days. This Steven Shielberg film from 2001 had a robot who is a boy named David who was run by AI. AI has been around for so long that in fact people feel it had lost relevance. This time, It is a little different. The fundamental change that has happened in the past few years is the computational power that is available to even individuals or small start ups that want to do complex work. Five or ten years ago, if you wanted to get access to a super computer grade computing of that size, if was almost impossible for you to do anything unless you are a billon dollar enterprise. Now, with pay per use computing and cloud, if had become very easy for people to really large computing infrastructure. So what has happened with that is low of the problems that are left untacked because they were just not been able to be turned around at a reasonable time have been have been picked up again. So, we have solutions that can crunch millions of images in a matter of a few minutes which used to take a long time with traditional computing. Once this computing power came not play, a lot of people started building really innovative solutions which demonstrated some fundamental learning behavior into new systems and that’s where most systems started showing some form of reasoning or learning and started calling themselves as AI applications. It is actually here for good this time. If you are a game of thrones fan, instead of Winter In coming, AI is coming is very relevant to technical people. There are signs of AI coming all around us. I will show you a few examples:
This year at the google IO, google changed its strategy from Mobile first to AI first. A few years ago, when google realized that mobile is become very prevalent and almost everyone everywhere has access to mobile applications, they had changed their strategy to mobile first, which meant that, everything that they did, they did with the focus of mobile first. There was a decision that they had to make internally that if there was an application that they had to make, A or B, they would make it mobile first which would give them an advantage on the mobile side. This year, they have switched from mobile first to AI first which goes to say a lot on what they think about AI.
If you look at Facebook, they have been investing on AI too. In fact, yesterday there was a news that they use AI to tap into users who might by potentially suicidal. And make sure they can reach out to them and make help available to them through algorithms.
There is another case that I find very fascinating. Apple, is known for their secrecy. Anything that they do, there are a few leaks on what they are developing and how its going to come out. But they made their exception on their secrecy for AI. They had realized that, when they started with CB etc, way ahead of competition, Their solution did not evolve as fast as others because they were not sharing what they were doing with a lot of other people to comment and offer their point of view.
These are four of the worlds most biggest consumer companies. But if you look at the enterprise side as well, lot of companies are catching up on the AI side. Here is a list of acquisitions that companies have made in the last year from CB Insights. You can see the acquisitions that have been made by companies like google etc, but companies like Ford and GE are also are trig to pick up a lot of AI stuff which is something that is impacting a lot of businesses across the board and is not just consumer companies.
All this is fine; Lot of consumer companies seem to be doing a lot of interesting stuff on AI, but what does it really mean for enterprise applications? I thought a good way of looking at AI’s applicability would be by breaking what does intelligence fundamentally mean. If you think of Intelligence, It is made up of five fundamental capabilities.
- Knowledge: Your capability to harvest knowledge.
- Reason: Your capability to reason.
- Perception: Your capability to perceive sights sound, touch, smell.
- Planning: Your capability to plan based on what ever you know.
- Natural Language: Your capability to understand and communicate in Natural Language.
Together, all these five capabilities, bring together intelligence. If you are able to replicate all these artificially, it brings you Artificial Intelligence. A machine which exhibits all of these traits together might be something that can pass security tests of humans vs machines. A lot of companies are doing work on either one or two of them or multiple of them depending on their use case. I will quickly walk you through what each of these things mean and what is happening with each of these capabilities. What I’ll also do is, give you one expanse or a reference story for each of these capabilities to help you understand these capabilities a little better. Some of them are stories of our customers and some of them are stories of other companies globally.
First, to start with knowledge, knowledge fundamentally means what you have learned in the past. Understanding whatever you have read, whatever you have seen and what ever you have experienced. In terms of system, it means looking at a lot of past historical data; whether its transactions, campaigns that you sent out whether its customers that visited your website and stuff like that and then trying to make sense of a lot of that big data or large data. Most companies use this fundamental capability to understand their customers, what are they doing, how are they doing and to predict what they might do next. You see a lot of these things manifested as recommendation engines. When you go to a website and buy a product, it recommends you something that you might want to buy along with that product or something that you might want to upgrade that product to. So the cross selling – upwelling happen based on whatever people like you have done on the website in the past. Demonstrates this fundamental learning behavior for enterprises to harvest knowledge to predict what might happen next. I’ll tell you a small customer story on how this is working for an enterprise. The particular company is a cable TV provider. Say, you and I were living next door to each other in the US and we had cable subscription from the same company. You and I are both watching, say, game of thrones. We get to see the same ads no matter what your demography in and my demography is. You might be a bachelor who is just out of college and it might be totally appropriate for you to see a certain kinds of ads and I have two kids at home and I might be appropriate for me to see diaper ads for example; even though we both are watching game of thrones. So, this is a customer that we helped by analyzing lot of consumer behavior on set top boxes to figure out what are their customers watching and from that, extrapolating to figure out what is their customers profile, how many kind of people, what kind of people are there in this household and using that information to segregate these customer so that appropriate ads can be shown to customers even though they are watching the same program next door to each other. The sis one of the examples of how knowledge in enterprises work.
The second capability is the ability of a machine to reason. What this is translates to is an algorithm evaluating pros and cons of a situation and then making a call on what to do. The fundamental thing is, once they make that call, also to learn and understand weather the decision was right or wrong. Lot of enterprises use this to make sense of images like xray scans or financial transactions or other automated systems where after going through lots of past data algorithms predict whats going to happen and then predict a course fo action based on that. Again here, one of the main things is to get a feedback on what ever the algorithms have classified. For instance if you look at a personalization engine, say you as a shopper go to an commerce site to search for a dress. Depending on what your persona is, dresses that are more appropriate for your demography like if you are a teenager then dresses mean something different to you. If you are office going professional then dresses mean something different to you. So depending on that the system tries to make a call on what you mean when you go to search for dresses. Again, the important feedback in this case might be when you search for dresses and your display renders dressed and you click on one of them and one time the system tries to understand that when I show you this result, this is the one that customer click on the most so it should actually become the first result. So you see a lot of reasoning capabilities in enterprise systems in the from of classification systems and even algorithmic trading where you have algorithms which trade stocks on your behalf based on the goals that you set are all based on their ability to reason and invest in multiple seconds.
One of the stories around this is a Fraud analysis for one of the companies that deals with financial transactions. Their financial transactions are usually based on employee transactions and based on all of their records of customers they had in the past, we built a solution that can predict which expenses look potentially fraudulent based on the customers behavior, the location where the behaviors has happened, what kind of company the guy belongs to, what is his role and things like that. So, this model fundamentally worked on algorithms capability on why It thinks something is suspicious and then trying to tag it out, again getting a feedback on whether it was right or wrong to further strengthen its data model and stuff like that.
The third capability is the ability to perceive. This is one of the capabilities that is under development. The computer vision capability is pretty awesome these days and even the touch capability is pretty good with apple’s 3D Touch and a lot of other devices that let you work with a touch screen. But the ability to smell and the ability to sense os still a work of a lot of academic research. But the way enterprises use perception capabilities is to eliminate tasks that need some basic cognitive behaviors from a human being. So say you were to go through a large volume of images and you just want to look at an image and say ‘Does this have a tree or not?’. Traditionally, you would have many people looking at these images and quickly clicking keynote fields. This is something which systems can do on their own. Without involving a human being. That is something we see a lot. Any Image recognition systems that you see are a part of the perception capability of AI. And one customer story that I want to talk about today is again from a Real estate company where they go through millions of images of houses everyday and the main thing they wanted to do it to understand these images a little more. So, the way it worked is, when they upload an image of a house for instance, we break to down and say: In this house we see these things like this is an image of a kitchen, this has got a range wood, It has stainless steel appliances, it has got a fridge, a dishwasher and stuff like that. Another thing that we did was, if you were on their portal and you happened to see an image of a house, and let’s say you did not like the price of this house or were to be too far away from your work but you really like the kitchen of this house and you really want to see other houses with a similar kitchen, then we would do that image comparison for you and pull out houses that have got similar kitchens. In this case, an island kitchen with a granite countertop and stainless steel appliances and stuff like that. Another thing that we could also do based out of this is detect property condition. If you get a lot of property’s images we can tell you that this peppery is in a pretty bad shape and you don’t have to spend any time analyzing this property in any more detail and just ignore this property for now. So these are all instances of computer vision capabilities that are being offered as AI solutions by many companies.
The fourth ability is the ability to plan. This is one of the most complicated AI capabilities and not something that is very prevalent today. The best example of autonomous systems that can plan is a GPS application of the Waze app. Depending on a lot f signals like traffic, road conditions, accidents on the road and stuff like that, it can navigate itself and say – don’t go straight, turn left, take a detour to reach your destinations faster, so, this is again something that works out of a lot of data for base learning and it evolves more as customers use it and it shows signs of intelligence. Where you see it is a self driving car. Very close to our office is the Waime headquarters where you see a lot of self driving cars driving around and trying to learn from traffic conditions and how people drive cars and automatically be able to drive cars.
The last capability and the one which is most widely used today is Natural Language Processing. So this is where you try to understand what people have written in large documents, emails, contracts, social media conversations and also derive conversations out of what you understood. So, Natural Language processing as less natural language invasion. Lot of enterprises use this today to automatically process feedback, to automatically service their customer support, to make sense of what customers are doing on their eCommerce websites for example. All the automated assistant like Siri, google or Alexa or even chat bots that you see are based out of understanding the intent of natural language query or conversation. So, example of something similar that we do for retailers is looking at product descriptions on their websites as well as what customers are saying about their products in either social media or as reviews in their websites or things like that and trying to make sense of what is important to customers when to comes to a particular product category. For instance, for shoes, what are the top 20 things that customers care about when they talk about does. So, we talk about our product discovery algorithms as as sales guy who has made tens and millions of sales conversations and it knows when a customer has walked into a store to look at shoes, what is it that the customers care about. For shoes, the number one thing that customers care about is where they are going to wear it and what they are going to do with it. Are they looking for shoes for a cocktail party, are they looking for a shoes for running or for what purpose and we try to automatically make a conversation based on what we have learnt from crunching these conversations. So this is one of the uses of NLP.
So from these five capabilities that you have seen, there are a couple of things that I wanted to point out. All of these capabilities depend on learning from existing data sets. So the learning behaviors shows itself by going through large volumes of past data, and to then do something or take some action as a human being would and that is kind of a fundamental solution for an AI Solution. The quality and quantity of data that is available for training data will decide how good or bad your AI implementations will be. And a lot of these examples that I have gives for each of these capabilities, they usually exploit one of the capabilities of AI. In my mind, that is more data driven intelligence than AI. AI system typically uses two or three of these capabilities together to exhibit sign of intelligence. When you mix two or three capabilities, the solution is more potent than just using one capability. For an example, if you were to talk about a retail customer that sells dresses online, the way an AI system or out system would work is; You upload an image of dress and then it will try to describe this dress automatically by looking at product description of other dresses, customer reviews that customers have placed on the dresses, as well as the computer vision part of it. So for instance for this dress it figure out that it is an off shoulder dress, it’s got half sleeves and it is something that customers wear for summer. This is all based on a combination of computer vision and natural language processing.
Another important case that is making a lot of come back is OCR systems. ORC technology is one of those technologies that has been around for a long time; but the main problem with OCR has been its accuracy. It has been really difficult to get 70% or beyond 70% accuracy form OCR systems. But with all these machine learning capabilities, what has happened is this problem is being looked at again.
For instance, a lot of the work we do on data extraction using AI technology uses a combination of deep learning to figure out which are the parts of data to extract information from, then doing OCR of that data and then trying to correct the output of an OCR based on what we have automatically learnt from different patterns that we see. So, somethings as complex as this PDF document; I’ve tried to show four pages. We tried to extract data out of that and tried to create a structured output so that it can be used for further processing. Something like this was not possible a few years ago by using just regular OCR techniques. The point that I’m trying to make is, there are a lot of cases where when you combine more than one AI capabilities together, you got a much more potent solution. If you look at the overall landscape of the AI use cases, they are all related to these five fundamental capabilities. I have tried to list some of them from recommendation engines to chat bots to self driving cars and stuff like that. This is by no means a comprehensive list. There are a lot more use cases; I don’t have space to fill all the use cases, but it gives you an idea of what are enterprises doing with each of these different capabilities.
What I wanted to talk next was, in the new year if you guys are planning to bring AI into your enterprise, where do you start? If you haven’t done any AI projects so far, how and where do you start and what are the few things to watch out for. For lot of our customers, we recommend first time AI explorers to start with an AI strategy workshop where we look at a lot of your existing systems and the kind of work that you do to figure out what is the right way for you to start your AI exercise. These are the three strategies that we start with:
One is the case for a customer who has a lot of data but is not really sure of what can come out of this data. So in that case, we start with a data exploration phase and say, hey because you have millions of records of your past customers, you can now automatically start to do raw detection based on that data or you have a lot of data on your customer’s purchase history, so you can do targeted advertising for them based on all the data. So that is one way of starting an AI initiative. The second way of starting is with a goal. You say, I want to cut down my operational cost by 5% and then we start looking at what is the opportunity to automate different things and different process that cost a lot of money and replace some people with algorithms that can do the work faster and in a more efficient manner. So that’s a new way of starting a new AI implementation. A lot times what customers also do is, they have a story. They know the use case that they want to solve. This could be something as simple as ‘I want to sell more products’ or ‘I want to do more cross sell-up sell’. They may have some data, they may not have some data, that might need to be procured from outside. But when they start with a story we figure out what does it take to execute their story in a more efficient manner. And within this exercise, we try to figure out whats a good AI strategy to meet the end goal for this customer. This typically lasts between one to two weeks depending on the size of the problems that we are trying to solve and it offers us a good direction to begin the AI process.
There are also a few things that I want to point out that you should bare in mind as you work through your AI implementation. So I had written this article some time ago on linkedIn with a title ‘Data is the new IP’ and I think companies like google figured it out way before most of the people that data in going to become the new oil or the new IP. These days there are a lot of algorithms or products available which can do pretty amazing things with Data. But what differentiates one companies IA with another companies AI is the quality of data that tech of them have to train the algorithms. So what I have advised most of our customers is, when you are looking at an AI solution, hold onto your data very dearly and try to get algorithms instead of data. Work with algorithms that bring their algorithms to your data rather than companies that take your data and try to do something with it.
The second thing that we have learnt from a lot of our AI implementations is not to discount the manual aspect especially during the initial phases. It is very important to factor that your first version of AI platform is very faulty; It is going to need a lot of hand holding and supervision. My advice to most of my customers is do not write claim for a fully autonomous system in the first row. Human intervention is now bad. In a lot of cases is it very important to have human intervention. The other thing you should also keep in mind is should you have enough training data to learn from. What is mean by that is, having data and having data that is good for training are two different things. For example, you might have a lot of record of the people who visited your website, left their phone number and name for contacting as a lead but very of them responded. You may have millions of these records but these records are not useful for training because they do not tell you anything about the customers. Other then the name and lead, they do not tell you anything about the customers that converted. So, it’s not about having a large volume of data, the data should have the right dimensions in it for algorithms to learn. That is also important when you start. The two things are again those that customers overlook is learning from feedback. Feedback loop is something that most customers miss out when they roll out a learning system. It is very important for a system to get a feedback on whatever predictions it has made whether it is right or wrong, whether a user clicked on it, whether a customer bought it or whether the accuracy is high or low for it to be able to refine whatever it has learnt from the data. So, make sure there is a feedback loop in what ever applications to roll out to begin with. Also, to think about what are the right places to start your AI implementations, usually what we have seen is these are things which need low cognitive capabilities such as things that you have in your enterprise that you want humans to see or respond. They’re actually good candidates to begin your AI and anything else that you have traditionally outsourced to a BPO for either data cleansing, data extraction, customer support, they are all good candidates to to begin your AI initiatives.
The other important thing that I can’t talk enough about is, incremental and controlled AI is the right way to go about things. It can be dangerous to build an AI system and to expose it to your customers directly without any supervision. And a case in time that most of you would have experienced are Chatbots. A lot of customers who roll out a chatbot to handle customer queries or customer support might have experienced that they backfire more than they actually help in automation. And the reason is, when a customer tries to talk to a chatbot, he is expecting that his problem will get solved and when it doesn’t get solved, it leads to much more frustration than gaining experience points on automation the experience. In fact, for most of our customers, what we recommend is when we roll out a chatbot implementation, we try to supervise the response from a chatbot during the first few weeks to understand if it is sending the responses right and if it is not, then we manually get somebody to fix the responses and we get the chat bot to learn from the fixing. But it is very dangerous to let aN AI chatbot in front of your customers. It can cost you your reputation, it can cost you a lot of your customers and a case that I wanted to point out here is this chat bot called ‘Tay’ which was released earlier this year. It was supposed to make casual conversations with people on twitter and try to imitate the talking experience of a millionaire. So what is did was, when it started interacting with people on twitter and other channels, it also started learning about how people talk about different things and because to was unsupervised, it went out of hand very quickly. It started swearing, it started saying a lot of wrong things. Algorithmically there was nothing wrong with it because it learnt from what people spoke with it. But unsupervised, it created a lot of bad press for Microsoft and they had to yank it back and stop it. This is, I think one of the big take aways for me that I would like to pass on.
The last part of this deck I wanted to talk about: If you want to start an AI implementation, how should it look like. We usually recommend a three phase plan. The first one is to have an NVP of your AI solution out. We recommend doing this within three to six months to begin with. Phase two is more about adding mobiles and business to whatever we learnt in phase one and a few more signals. And, phase three is where you try to build the autonomous learning system all on its own. Typically the activities you should aim for in your phase one is do an NVP which solves three to five features or signals that it analyses and try to have a reasonable measure of success or failure of this platform. It should not be binary. It can not be a hundred percent success or a hundred percent failure. But try to make upfront what would make a successful AI implementation for you. Plan for human training and supervision. Do not let your AI system in the first phase run rogue. Try to use a single capability if you are doing an AI solution for the first time. If you are using Natural Language, try to use a solution that uses only Natural Language. If you are using computer vision then just use computer vision. For a first time, it a bad idea to combine more than one capabilities in a single use case. It’s just complicates everything and takes a much longer time to get anything out of it. But once you have successfully implemented your phase two, then you can add another capsule of 6 to 12 months of refinements on top of that which essentially would augment whatever you did with phase one with external data sources with data from outside your enterprise. Add more than one AI capability. So if you have Natural Language and now you want to bring imaging or you started Imaging and now you want to bring Natural Language, that is a good idea to do in phase two. Again in this phase also, you should have some human intervention and supervision but a lot reduced from phase one. But, you can add a lot more signals in second phase and also more importantly also learn from what ever you received in the feedback loop. If the predictions you made were not right or the conversation you made were not right for the conversation you made were not effective, try to figure out what went wrong and try to fix your algorithms or data models accordingly. In the final phase, once you have done that, that is where you should explore unsupervised machine learning for automatically detecting anything that the system might have come across that you you might not have taught it to. On phases one and two, it as also a good idea to keep running your legacy systems. For example, if you have your BPO or back office supporting customer support ticket, to begin with you might have to give just 10% of your tickets to the AI systems and keep your existing system serving 90% of your tickets and slowly cut it over. By phase three you should be in a position to make a call whether you are completely going to cut it out or not. It should still have human supervision is what I would say as the system is still learning and sometimes it can learn something wrong so there is no harm in supervising while it is learning and it usually pays off by making that investment.
Th last slice that I wanted to talk about was this championship that happened a year ago in march. For those of you who might not know this, google acquired this company called Deep Mind and they train their algorithm to play a game of GO which is supposed to be much more complicated than a game of chess and they played with the world champion Desarol. Th algorithms won the game 4:1. Out of 5 games, if won 4 out of them and at that time there was a lot of press about how this is actually a big question mark on the capabilities of AI system going forward and what has this got for us as humans and stuff like that. And some of the moves that the algorithms made, game experts call them as Devine moves like the “Hand of God” kick from Maradona. So from what ever we have seen and from whatever our customers have experienced, we can definitely see that AI is going to fundamentally change a lot of ways enterprises do business today. In my 20 years of career, I have come across two fundamental shifts that I remember. The first one was the coming of the internet which fundamentally changed how a lot of customers sold their products or services or how they interacted with customers and stuff like that. The second shit came with mobile revolution where everybody n the world had mobile phones and that also fundamentally changed how people make payments, how people interacted with each other and stuff like that. But both these revolutions also packed bucket loads of opportunities and I think AI is again that things which packs a lot of opportunities in it and I hope you have a good 2018 planning to get those capabilities into your system and I hope you learnt a little bit about AI through this webinar. Thank you.