Im gon na go ahead and get started then well welcome everyone to the seminar. My name is Kevin. Dewalt then jump right into things. Tell you a little bit about my background and while were here first off, my name is Kevin and I am co-founder of pro Lego, along with my business partner, Russ Franz, and so, if you hear me mention the name Russ today, that’s that’s, what Im referring to? We started the company a couple years ago to become really the boutique AI service provider.
To large enterprise companies, so we spent our days helping banks and insurance companies and telecom and automobile learn how to adopt the high technology. But as first my own background goes, I recognize that a lot of you here are entrepreneurs or may be looking for a small company product ideas, and so Ive been doing this for about thirty years. I did my graduate work and when I was at Stanford in the early 1990s in neural networks, right during the height of the
Ai winter and several decades later, I can finally put all of that path to good use. In my day, job but Ive been in tech and starts. My whole career been a founder. A number of times worked as a, venture capitalist. Im still an active angel investor and Ive done all, kinds of large infrastructure tech. Data processing work in large companies um you know so the biggest of the big. That are always looking for a competitive edge and intelligence financial services on Telecom, the usual suspects. So I dont. I would imagine that most of the
challenges in quest you have I can probably answer if I cant answer them during the presentation I will do my best to do so during the webinar or start during the Q&A afterwards okay so were paying a little context for what you can expect today when Russ and I started the company two years ago one of our missions was to help people understand what AI is and what we can do with the technology and we look around the universe we find that you know mulleted the information out there about.
a tie falls into one or two buckets its other high-level sort of marketing fluff you know you know robots you know automating stuff that’s basically functionally useless or its so detailed and technical like you know taking online tensorflow course that for your business person or your entrepreneur it doesnt really help you answer two questions and that is what is this and what can I do to it and so that’s were gonna try to deliver today for.
context if you get lost if you’re unclear if you want more details everything Im gonna talk about today is captured in her book its called become an AI company in ninety days and you can download a free e copy from our website the URL is up there its at our website pro Lego dot o we think its really the first practical business book on AI so whenever I talk about tech to anyone Im rather than just kind of bladder.
information that you I want to give you very specific skills you can use things you could take and apply to your job right away and so by the end of this twenty five minutes I want you to have two skills one is it I want you to understand AI basics so when you go into a meeting and somebody talks about machine learning or training data you’re gonna know what theyre talking about and then and so if you’re trying to recruit an engineer or you’re talking to.
a potential customer and heres somebody in the room say something like in q2 or gonna apply machine learning to our data Lake you’re gonna know that that person doesnt know what theyre talking about so were gonna help explain what those words mean and how can you apply them in the second Im going to give you some specific tricks you can use this is the.
stuff that I actually use in my day job with my clients to identify opportunities for using AI so sort of simple product patterns you can use look for new business opportunities and new product ideas okay so lets jump right into it okay part one lets talk about some of the fundamental ai ai fundamentals what do all of these or its me.
so first if you are if you’re not if you’re confused about what the term AI means you’re not alone I find this as employing the most hackneyed confusing terms in the industry and Ill just explain to how I use it AI just means intelligent computers if you want to call your pocket calculator AI go right ahead and do so its just a general term I used to talk about the state of the industry and where we are at in time so just general time AI use means Thinking Machines machine learning.
is a type of AI and that’s why on your screen you see that its the size of n diagram and its a different way programming computers and machine learnings are really important turn Im going to explain to exactly what that means and finally an attorney may here is deep learning deep learning is a type of machine learning and it represents the state-of-the-art AI so to summarize.
AI general term just means thinking computers machine learning a specific way of building software and deep learning is a type of machine learning okay so lets talk about machine learning but before we talk about what machine learning is lets talk about what machine learning is it so 99 percent point 999 of all software that you’re ever gonna see is built like this with a developer or a programmer explicit explicitly telling a computer how to perform a task and so Ive got some simple pseudo code on screen here.
this is not a language is an example of some type of code and if youve had any kind of programming background you can kind of look at this and say okay I see with this you know this code is supposed to do its gonna print out the time once a minute for ten minutes right so runs through a loop one through ten prints.
out the time so my question for you is how did the developer the person that wrote this software know what to type like why didnt developer know that it was supposed to go from the mental 10 and not 1 to 11 and why is the system waiting every minute and not every 30 seconds or every 59 seconds how did the programmer know what to do and the answer is somebody told her how to do it.
right a human being whether that was a customer the boss a product manager maybe was a specification maybe its information they learned in a log file but somehow they got information from an external agent that described a problem and then when they went in them and told the computer how to solve the problem so the developer looked at the world and developed this abstract language we call programming which it used to instruct.
the computer okay I know Im belaboring this point but its a critical topic because machine learning doesnt work like this in machine is it machine living its a different way of programming computers and the machine learning a developer uses data to training model how to perform a specific task were going to come back this definition a lot so this might seem a little abstract to you if you’re not familiar with these terms lets start with an example lets pretend for a minute youve got a new startup idea or.
a new product the idea or your service company trying to provide a no software services to an industry and lets pretend for a moment that that industry is a bunch of scientists who are trying to detect earthquakes and so they’ve got a bunch of sensors scattered all over the United States maybe have fault lines maybe in the Midwest and those centers gather signals and the scientists use those signals to try to predict an earthquake so what are those sensors gathered things like temperature.
pressure vibration time of day humidity any information that might be predictive of an earthquake and all that information is coming in to the scientists when you can imagine that the scientists are after gather this information dont want to look at every signal from every sensor that would take too long and be too tedious so theyre looking for you to design an intelligent signal alert system a system that would sift through all of those signals and.
only notify the analysts of the scientists when they need to pay attention to it a signal because it might indicate a potential earthquake so lets start pretending like were going to build this system as as a traditional software development project so what you would probably do as a developer is sit down with a scientist and say hey tell me what makes an interesting signal and they might tell you something like well if its in an important region its interesting for instance if its in the.
Sun Andreas Fault in California really care about it if its in the middle of Birmingham Burnie Birmingham Alabama which doesnt have earthquakes I probably dont if these sensors and maintenance I dont want to know about it if the vibration starts picking up maybe I do want to know about if there’s additional pressure on the sensor may dont want to know about it I know what the rules are but the scientists would probably describe what they want and then the program would go on and build software that tries to tries to apply that logic.
than what the scientists want to my signals that might be interesting to the scientists okay pretty straightforward basic programming right here the problem with this kind of software is that it can get pretty complicated as you have so many exceptions and if you want to know that its sensors in this region but not at this time and you want to know that.
there’s a certain amount of vibration but only if its followed by additional vibration as the logic gets more and more complicated this software can get really really complex it gets hard to maintain it gets brittle it doesnt adapt as well it doesnt that the changes and it doesnt work as well in long run so it gets more expensive it doesnt work as well so that’s why a lot of companies are increasingly using machine learning as an alternative to building software like this so again lets go back to our definition in.
machine learning a developer uses data to train a model how to perform a specific task so if we were gonna build our intelligence in an alert system using machine learning step one would be to go and gather the data we would sit down with a scientists or maybe their IT department and we would say hey lets put together two spreadsheets one spreadsheet has all of the interesting signals and another spreadsheet has all of the not so interesting signals and.
were gonna split them up into two different folders and once we have that data and machine learning engineer is going to take it and feed it into an algorithm or a model that’s going to generate predictions and were gonna feed each one of these signals into the algorithm and its going to try to predict is it interesting or its not not interesting and every time it makes a prediction he gets it right or wrong were gonna give that feedback back to the algorithm so if we start showing this algorithm enough examples eventually its gonna get good enough to.
the point where its able to predict whether a signal is interesting or not and that is machine learn if you get nothing else out of my discussion today an understanding of what machine learning is and isnt is viable career advice you’re gonna be able to apply in the next decade I hear too often people using this term out of without really understanding what it means but it really is just a different.
way of building software so to summarize the traditional way of building a system like this would be the programmer explicitly telling a computer what to do but in machine learning we take examples we feed it to an algorithm and we train the algorithm how to know how to do something in this case make a prediction that’s machine learning ok so lets dive into deep learning so deep learning is state as I mentioned before is a type of machine learning and its also the state of the art in AI and.
I think its fair to say that the only reason were talking about the AI today is because of deep learning if you want a really easy read sort of explanation for why this is happening now and what the big picture is I highly recommend this article in The New York Times Magazine its about a year and a half of all that its still excellent and very relevant its called the great AI Awakening and its about the history of what happened at Google and neural networks and how they started using it.
for machine translation so an excellent background on the technology if you’re interested ok so lets talk about deep learning what makes machine learning deep learning so a deep learning deep learning system usually is comprised of three things starting on the right hand side first is a neural network that’s a type of machine learning algorithm neural networks have been around since the 1950s theyre not new they tend to be a lot more complicated than traditional machine learning algorithms.
but they tend to generalize to big complex problems much better than other models the second attribute is large data sets so using lots and lots of data to train these big complex neural networks and because we have these big complex models only large data sets we have to train them on a specialized type of hardware called the GPU of or graphical computing unit the number one manufacturer of which is Nvidia which is.
why Nvidia stock has been doing so hard over the past couple years so deep learning has three components neural networks lots of data and GPUs but so what right why do we care what can we do with deep learning why is it so special that’s two reasons why first off deep learning is helping us solve some really really hard computing science problems the things that MIT researchers have been.
working on for decades and so when people say deep learning changes everything what they mean is there has been a rapid explosion in solving hard fundamental problems like computer vision which were going to talk about in part two or natural language processing which well also talk about so very very hard computer science programs that next researchers for decades well in the past couple of years.
researchers using deep learning have made faster strides on both of these problems than all of human history before and that’s one of the reasons why things are so exciting and you can see this time to evolve into your products if you have the iPhone 10 youve probably noticed how great the biometrics is the facial recognition when you hold it up to your face that’s all based on computer vision and if.
youve been speaking to your come up you to your phone in the last year you probably knows that it gets better and better at natural image processing that is taking your spoken words and converting them into text okay so deep learning leads to fundamental breakthroughs the second reason is that deep learning is great at handling handling complexity so anyone uses this graph but I think illustrates is punt really well on the y-axis we have performance and on the x-axis we have.
data and so this graph measures the trade-off between the amount of data you have and your performance and so when the human being with people we do really good at small amounts of data if you show a small child a picture of a zebra you only have to show her two or three examples before she understands what a zebra is but as keep adding more and more data we dont tend to get any better in fact we sometimes get overwhelmed we get worse traditional machine learning models that is things that are not deep learning tend to go.
tend to improve very rapidly as you more data add more data but eventually they kind of level out and stop getting better right so they might achieve human-level performance are a little bit better better but then they sort of peter out with deep learning you can theoretically keep adding more and more complexity adding additional neural networks additional layers more complex algorithms more data more computing power and you can keep getting better and that’s why when we talk about the power of AI and why its such a fundamental technology and why its.
gonna change everything this chart is why its because we can solve hard and harder problems with this foundational technology okay so going back to our intelligence in a alert system if you want to figure out ways we could add deep learning to the system and make them more accurate you could take images of wraps from a date you know from a drone or from a satellite and feed them into your machine learning algorithm to make better predictions as.
to whether a signal is interesting or not so deep learning allows you to add new data sources that you couldnt traditionally do okay so weve talked about that the basics this couple of the terms I want you to understand so once again going back to our definition of machine learning on machine learning a developer uses data to train a model how to perform a specific so lets talk about data letters I mentioned before at the beginning that when we talked about building artists intelligent signal alert system that we needed to gather a.
bunch of training data and were gonna get some you’re gonna sit down to the scientist or the analyst and put all of the interesting signals into one spreadsheet and all theyre not interesting signals and another spreadsheet so if you were paying attention you probably thought yourself like that sounds like a lot of work and you’re right it is and in fact acquiring.
and building training data is generally the most expensive highest-risk part of any AI initiative so if you are an AI entrepreneur if you’re doing an AI innovation project within your company one of the first questions you’re going to ask is how am I going to get access to a predictable unique data source to do my product that’s also one of the.
reasons why I started a services company working for large fortune 500 companies because they have data assets you need to do this type of work okay so once we have on data did it consist of two components it consists of inputs the the information that you feed to your machine and it consists and outputs the things the marbles are going to predict and so if you’re going to talk about any AI strategy the first thing you want to ask.
is what is my output I cant tell you the number of times I talk to an executive an entrepreneur they talk about doing machine learning on this side of that project and I have to ask them what are you trying to do what are you trying to predict mean machine learning is not some magic wand people wave it around that it solves all your problems so the first step you want to do when you’re talking about a machine learning.
product or project is what are we trying to predict are we trying to classify document were trying to predict the future cost resource are we trying to figure out wheres an image are we gonna predict the probability of the event we gonna summarize a document are we gonna recommend a product to a customer are we going to generate a probability event for a cyber intrusion is it or we can.
identify a spoken language so whatever your business goal is not start your strategy or product I think about what am I trying to do what is the output the same thing you want to do is think about do I have training data do I have inputs that are going to be predictive of the output and this is often going to be one of your hardest most expensive questions to answer so what are the inputs well if.
its a computer vision problem its the pixels of the digital image if it is a natural language process it is the transcribed text from written or spoken speech it could be time timestamp events from sensors historical sales records the prices of homes sold in your area that you know the the possibilities are unlimited but you have to identify something that’s gonna have some.
predictive power for what you wanted to achieve okay the next item apologize if you see a cat running across here my cats wandering around the office at the moment the next item in our definition is models and so models or algorithms are the the code that you train with your data to try to make predictions and so Im going to give you the an example of the worlds simplest model so pretend for a second that weve got some training data the inputs are on the x axis along the bottom of your screen okay and the y axis on the output and.
the the training data you had was this scatter plot on the screen or its listed in the columns on the right-hand side of your screen so youve got X Y the input we got output we want to build a machine learning model that’s gonna predict Y when we input an x well as it turned out you had it you learned a perfectly good why dont your eighth grade algebra class and that is y equals MX plus B which you may recognize is the.
equation for a straight line so in this particular instance are you going to try to predict quite out the which is why based on our input X and were going to try to discover when we train the model what is M and what is B and were going to feed these examples to the model to try to find the correct and then B and of course you dont really need machine learning in to solve this problem you can do it as a regression analysis in Excel but for the sake of this example.
why are M is equal to 0 1 and b is equal to 0 5 and that’s the equation for the line and that’s the best that’s the best solution that this machine learning algorithm can come up with you can see a bug on your screen and that is really all a machine learning model is and if you’re interested this is also the worlds simplest neural network it just.
one sliver of a neuron than a neural network so a bit of trivia there in case you want to impress your friends at the cocktail party okay some common models you may run into and for us multi-level multi-layer perceptrons convolutional neural networks recurrent neural networks really dont get caught up in models the stuff is changing all the time there’s tons of them off the shelf you want to start your project by using off the shuttle or wherever you can and researchers are coming up with hundreds and hundreds of new ones every single day and Im saying that with that.
exaggeration okay so that wraps up our basics we talked a little bit about the fundamentals of AI and what these terms mean so now for the good stuff what can you do and I know from some of the advance questions a couple you have questions about well how can I apply AI in my business Im gonna give you some tools that you can use to figure that out so you know this is the trillion dollar question what can you do with AI and if youve tried to figure that out.
youve probably run into a lot of technical right understanding what a convolutional neural network is and what it can do or an lsdm or it doesnt really help you answer the question so rather than talk about this low-level tech weve developed a constantly called product patterns or AI product patterns and these AI product patterns are generalized ways of talking about what the technology can do is apply to a common business problem and so you can talk about what a convolutional neural network is which I have a little example here on the screen.
but that doesnt help you very that doesnt help you very much but if you know what computer vision is you can think about how to apply that in banking or travel or a healthcare or for online sites and so were going to talk about for AI product patterns you can use these four patterns as building blocks for identifying potential startup opportunities or opportunities in your company I use these product patters on every strategic in consulting engagement I have with my clients when they start talking about a new idea the first thing.
that goes off my mind is what product pattern are we talking about so lets talk the first one the first one is computer vision computer vision is a set of algorithmic techniques to learn about whats in an image and try to figure out so computer vision is as I mentioned before when we talk about deep learning is what is problems in computer science.
that people been working on for four generations right and they’ve made incremental slow progress every single year well in 2013 a couple of researchers decided to throw a very large deep learning convolutional neural networks at the problems and in 18 months some very very hard problems or instantly solved one of the common computer science problems you may have seen is called cats versus dogs where you just get a set of cat and dog images and you have to build an algorithm that will predict does this predict this is a.
picture contain a cat or a dog well before deep learning the best research could get about 80% accuracy with deep learning this has now become a trivial task and in about 10 minutes using off-the-shelf tools you can build an application that will predict a cat or a dog an image with about 98 percent better accurate accuracy so really a computer vision as right now is the most.
advanced form of AI and this is as applying a deep learning and its one of the first spots you wont want to look for near-term product opportunities so what can you do with computer vision well you can classify images right does this image have a cat or a dog you know is is this you know is is this someone that’s breaking into my house is this somebody open the front door you can identify an object image like what is.
this what is this object and how do I describe it you can do image search retrieve a specific and Mitch from a large data set you can do image restoration so we did one small project with an automotive company where they wanted to have for your cars in the website with the computer vision you can take a low-res car image and turn it into a high-res car image its also a.
biometrics facial recognition so anytime you any type of problem we are interacting with creating or using images is a computer vision problem and to just give you a sense of how fast this technology is evolving and and how smart it is I like to use this little example from Google research on this is a year year and a half old so its getting a little bit dated and Im sure they’ve got a lot more advanced stuff now but I think it illustrates the point in this particular instance Google is trying to identify does this picture or.
try to look at this picture of a beach and identify which of the parts of the picture are people and which are kites and you can see in some cases this isnt this is pretty obvious on the lower left-hand side of the picture you know the algorithm is able to predict with 99% certainty that these are people and then the very top center is about 99% certain that that thing is the kite right so some of these are easy but if.
you take a look at the far right hand side your screen you see there’s a kite with 96% and a person with 87% and you can see these are really really tiny just little blobs especially the little black blob of the person so how did the algorithm know that that’s a person and that’s a kite from just a couple of pixels all right like how does it know.
how to do that and the answer is it learned from context that kites float in the air and people are in water but you dont see people in the air and you dont typically see kites in water and so that’s I just want to give you a sense of how fast technology is evolving and the kind of things you can do with it okay so if you want look for opportunities and computer vision if you have any application any system that’s already using images or you think you can use images of course.
that’s a great opportunity but the one that I see more often is a company or a system a process where you have an existing workflow and you come up with ways to add images to this but to pull you know Instagram feed or or you know our product pictures and add it to the workflow to make a better prediction you know drawing images satellite images if you have an existing prediction you can add images to this you can process them.
in a way in which you which was too cost-prohibitive before when you had to get people to look at every image training data is a challenge with computer vision if you want to predict whats in an image for example this is this collage of faces here if you want to predict the gender of some meat or if.
someones going head gear or if the person is wearing glasses but you’re gonna need to develop a set of training data and that’s going to require hand a person and most likely to go in hand label everyone these image and identify yes glasses no glasses with every image that you can make that prediction um this training data and labeling data is getting easier and easier and there are an increasingly number of third-party companies who will do this for you are there as a product or service okay so.
that is computer vision lets talk a little bit about our second product patent and the one that honestly I think its probably the if you’re an entrepreneur looking for opportunity I would say this is the one that’s most exciting at the moment because were really it is changing so fast natural language processing is a computing technique that interacts with you know natural language whats natural language its language created or read by people okay its not computer language its its English text Chinese text its its.
naturalism the kind of stuff we use every day the words that Im speaking right now so some examples of natural language processing our machine translation you know converting from one language to another okay Google Translate you see you awesome examples of this speech recognition you can see that in your right off your phone and it recognizes your speech and it translates that into text speech generation and you have an automated call center and you want to try to generate speech based on some background to the topic at snatch.
language processing entity recognition or extraction if you have a group of text and you want to plot a specific word or the most important word that’s a extraction text generation is exactly what it sounds generating text if you want to automatically right you want to write it in play if you want to write a poem you will see there are examples of people if you want to want to write a a stock report based on an earnings call their companies out there using NLP techniques in this kind of text generation the text summarization whats.
in this email you dont wanna read all your emails D&I no I dont recall my else a text summarization could only do that on text categorization is this email spam or is it or is it just you know email from somebody whos not a native English chatbots and sentiment analysis these have been out there for a couple years so those are some examples NLP but the reason I say that its most exciting is that natural language processing is about three years behind computer vision and it really didnt start the researchers didnt start using.
deep learning has sinned in the rfmp until about 18 months ago but its rapidly becoming sterile yard and if you look at all the work that’s being done on NLP now you’re seeing more and more deep learning come into most papers probably some end up using some sort of neural network some sort of deep learning technique solve harder and harder problems okay so how to spot NLP.
opportunities so obviously if you have an existing an LP solution you want to use deep learning I think this goes that saying in fact if you’re not already doing it you’re probably about to lose the competitors document categorization or summarization you know whats in this document where should we brought this report with this piece of paper so we take this analyst that Alice you know or Chinese north and what how we triage information college applications I remove a few college applicants is this a highly.
qualified candidate or a low fall or can that were probably not gonna be interesting whats the most interesting information in this email thread you know how often if youve gone into a thread on Gmail and read through you know email after email if theyll think oh my god well have to pay attention this thing well you just NLP whether the far the most interesting you know the most interesting piece of information and a couple of our identification and one that we are hearing more and more in.
financial services spaces doesnt seem you know does this memo violate any of our compliance processes for policies training data for NLP is harder than computer vision its easy for a person to look at a picture and you know say oh this is a cat or a dog or Christmas classes but to read a document to understand whats in it and try to.
classify it it tends to be a harder problem so it just takes more time additionally if you think about just the information density a 500 by 500 image which is pre low res image contains 750,000 data points or numbers but whereas if you look at a a single sheet of paper it has 4 or 500 words and so you need just a lot more natural.
language data to train a a problem than you do with computers additionally well I say I should say so really winning the game in terms of the training deal with NLP is often a task of redefine the problem and as an entrepreneur this is bringing more with customers to figure out you know whats the minimum quality spend that they would want coming out of.
them a machine learning algorithm it uses NLP and to give me some examples there’s a pic that I read that was trying to do some text classification so sort of with a training set from Wikipedia it took five hundred and sixty thousand Wikipedia examples and tried to categorize them into the one of fourteen top-level categories of course Wikipedia pages are well written you know they.
tend have a lot of jargon and so on the computer found it fairly easy to do this task was able through with 98% accuracy Dave sounds great it sounds like computer vision is everything but then if you look at the same people had some studies done on Amazon product reviews and in particular if you try to get an algorithm to predict one to five stars whether a the review of an Amazon product they can only do something about six percent accuracy but if you.
redefined that problem and try to predict is this review positive or is it review negative you get about 95 percent electricity so if you’re thinking about doing any kind of an LP problem dont and start by imagining all the great things you can do for the future but try to imagine the minimum success criteria that you can bring to the products.
that’s going to simplify crate-trained it okay so nine minutes left Im gonna go through the third product pattern and one of the most common ones its called next in sequence predictions youll see in the literature and training courses sometimes they call this tabular data but its basically trying to make predictions of based on information that’s captured in your databases or or tables or spreadsheets okay and were.
talking about things like you know information you gather from IOT devices or sales and marketing data or online user behavior where they go or they click on whether they look at or server longer things that are tabular that we that you can pretty commonly see in the database you dont hear actually a lot nobody gets a PhD you for for actually publishing a paper on next in sequence applications and that’s why you dont hear a lot about it but really I find.
that this is often one of the most useful tools you can apply in any kind of business setting because there’s so much information captured in databases and quite honestly its its often easy to identify what you want to predict and there’s a lot more training data so very common business opportunity even if the the cool kids working on a ai research dont focus on so some examples what you can do with mix and sequence predicting future sales classifying log entries is this a cyber attack a system failure or is this there.
is a machine about the crash because its preheating identify the fraud credit card flaws trying to predict whos gonna buy I got a website I got people coming on and I use machine learning to predict okay based on this persons behavior theyre likely to buy and try to like by that product right over there very common use of next and sequence.
so I mentioned structured data and so let me spend a moment on that for those that you dont have a computer science back background Im talking about the climate data you would expect to be stored in the database although you could store a blob of text like a like a tweet or a document in the database typically that’s not what you do usually well see those on a file system Im talking about things that are category categorical or continuous fields in the database so what are categorical fields.
things that have a finite set of a finite set of examples like store you know is it rest in Chicago or San Diego its got to find that number or continuous data numbers that can go from you know negative infinity to positive that have an infinite number of values like sales or temperature or if you’re using dance as a UNIX timestamp its time-stamped in seconds right all of this is is categorical data Im sorry is tabular data so if you want for next in sequence opportunities it really is its usually where you want to start is look.
at your existing data look at your existing processes look at your KPI so what is your organization accountable for you trying to lose sales are you trying to increase the the workflow of a particular operation of you are you trying to you know accelerate the closing of a case do you want to close more cases faster um it will tip me look around an organization if you look at the tape guys and what theyre measuring you will often be able to identify mix opportunities and usually you can also.
improve those by finding new third party data sources to add to any existing business process so instead of trying to predict sales based on past sell behavior you can also add in the weather all right look at the weather for a particular region figure out how people probably shot moments and I said sunny out in the dont shop as much when its snowy in and that with it okay training data is a bit of a different problem.
next in sequence so you rather than having to go and crate new training then you probably have plenty of data in your database training data actually becomes an engineering problem where your your data scientists are gonna have to go into the database and theyre gonna have to do is called feature engineering which is looking at your data and trying to derive features.
or inputs that are going to be predictive of your output for example you might have a time stamp associated with with every sale but you can take that time stamp and you can convert that into a feature which says day of the week and that way your algorithms can more easily identify trends in the data by pulling out specific features that might be more predictive okay so weve got our four minutes left here before well start taking questions and Ill leave plenty of time for Q&A so the.
final pattern is and the one that we see is one of the least is collaborative filter a collaborative filter is used when you have effectively in matrix of users and items and so the most common use of a collaborative filter is a recommendation engine so if you have an e-commerce site or you have self transactions you’re often going to want to try to predict you know what is this person likely to buy tastes past based on past behavior we can look at their.
past behavior but if theyre relatively new customers that might not tell you a lot instead a collaborative filter is an algorithm that looks at your past behavior and looks at all the past behavior of people like you and tries to make predictions on what you might like so the example I love to get it is myself and my network and Netflix so when its a Friday night and my wife and I wanted to decide to sing to figure out what to watch you know we really open up Netflix and the first thing I start.
doing is throwing a temper tantrum and saying no period pieces and no rom-coms it I dont wanna watch the silly wrong come but my wife has figured out that if the wrong calm has been someone or Ben Stiller or one of the actors that I like I will be excited and see it so you want it that’s not an easy type of that’s not an easy type of logic for algorithms to.
to generalize and figure out but I would bet if you love to cross all the different married couples and look at all a different all the different men type of rom-coms likely movies they like I went benching from a cluster of people who like rom-coms and Vince Vaughns and those people probably have similar preferences to mine and you can make a product recommendation that that’s a cloud a filter a user comes usually seeing those and like advertising campaigns or product recommendations and inside were talking about these these four product.
patterns and this really is the first the first filter that you can use when looking for AI opportunities of course this does not cover the realm of all AI research there’s reinforcement learning you never see Romars a lot of stuff that’s happened on the cutting edge and we do some of that with our clients but this is a good first triage and if you’re new to this topic and you want to find out for charities these for product.
titles can be really useful for looking for business opportunities final point Ill say is that with AI your own networks and deep learning is so amazing you can actually combine these patterns and so the inputs of these models are just a series of numbers and the outputs are also series of numbers and you can connect these different patterns together almost like Lego blocks it really is astounding if you come from a computer science background and Im amazed by this every day but you can.
build more complex alone networks and successfully train than using off-the-shelf tools by combining patterns using some of these techniques okay so final step Ive got a quiz to go through which we can take but before we go to the quiz Im gonna leave plenty of time for questions looks like there’s a lot in the Q&A here so were gonna skip the quiz our book has some quizzes in it if you want to brush up on your knowledge let me just finish up there and say thank you very much for your.
time if you want to talk to us more my best recommendation is to contact my business partner Ross friends you can contact him and connect with him in LinkedIn hes at russet or Lego that I owe if you connect with me on LinkedIn I know there are hundreds of you on this call on this call I will not recognize you we just mentioned try to connect.
with me that you took the East seminar from MIT said I can connect with you thank you very much lets take some questions why do you want me to jump in and start these from the top or do you have specific ones you want me to call out yeah that would be great if you just start as you see fit okay so from anonymous attendee which AI sub field between machine learning NLP robotics will truly have the biggest impact on businesses well so I will say that let me repeat the question which.
sub AI subfield between machine learning NLP robotics will truly have the biggest impact on businesses and of course if you were paying attention I dont know how to answer this question because I dont really understand it NLP is a is a is an AI product problem as we described it which is a a a type of solution that can be solved with machine learning so NLP for the most part is machine and a.
lot of robotics applications use computer vision and machine learning so Im not sure quite how to answer that and if I can predict which ones would make the biggest impact I would be investing in those companies and not telling you anonymous attendees okay so how should data be collected and structured to generate good insights from machine learning models what kind of machine learning models are available in what.
situations should they be used ok so I try to answer that with our product patterns so I guess what I can say is that the data really depends on what you’re trying to do so let me ask this question differently if you if you have an idea of what you want to do is they are the first thing you want to do is figure out what is my output what am I trying to do what is the what am I trying to create with this you know what.
am I trying to accomplish with this information am I trying to predict is in the in the bit in the whos in the picture Oh am I gonna try to predict which is next week sales the next thing you want to figure out do I have inputs do I have data that might be predictive or whatever that how it is and if you think that answer is yes then you might be on your way to finding a an opportunity for using a and which type of machine learning.
models are available there are infinite there to give you a space of how fast this is changing there are on archive which is a print publication site there are 100 papers published every day on machine learning I mean that’s effectively a conference every single day on machine learning and if even if only 10% of those papers are relevant and having new and innovative models.
that’s more than I have time to read oh there are infinite number of models and theyre ones coming in all the time you dont really need to worry about that you need to focus on what you’re trying to solve and if you have inputs like okay so since getting good training data is an impediment to supervised methods or unsupervised methods that dont require training data not yet good enough to compare and the quality of.
prediction with supervised methods I need to be always do we always think training data this is this is a good question so for the most part if you look what people are doing with unsupervised and unsupervised machine learning is theyre typically taking a set of data and trying to cluster it or organize it around a particular around a you know a particular in a set of values that theyre looking for you that’s typically where you see mostly unsupervised stuff these terms like unsupervised and supervised learning when you get into the details and you look at the crowd to try to solve they.
theyre never that simple and even people who are doing these things that are cortical unsupervised a lot of times semi-supervised nodes they have some training data and theyre supplementing with other data in some way and in some domains like were working on some reinforcement learning projects right now people think of reinforcement learning is totally different than supervised learning um but I can promise you that were having to generate exam on the project we have were having to generate training examples and billing and organizing that this still remains one of our biggest challenges Im so for.
right now dan still Kenny I know we may reach a time when you dont need as much but as far as like practical business problems most know most of you right now I see you’re supervised learning is still being needed what kind of off the shelf tools are currently used in the industry for computer vision and AI solutions so there if you just so if you if you go into any of the common platforms high.
towards tensorflow Quiroz fast the eye you’re going to see an example of the kind of office shell tool for the most part everything is and everything is sort of off the shelf in that when a researcher discovers something new and algorithm there’s a tremendous amount of industry pressure published right now because the researchers need to get feedback from colleagues to figure out if theyre really on something and so.
most of the most innovative techniques theyre out there people got published and so its off the shelf enough if you go to a segment called papers with code you can find the code you can use to to achieve the same results is what the researchers have off the shopping like an Amazon AWS and you can find you know computer vision algorithms like vgg 16 which is five or six years old then you can use to make image classification techniques and there’s off-the-shelf you know NLP applications that you can get read of readily and most of the next in.
sequence of pressure models you can get out of the library called cycle its a lot of this is out there the bigger question is organizing the training data and getting it pulled into a way you can use it so again I would spend less time thinking about models because models are always changing and if there’s a lot of fuss amorphous tough ones think about data what about for developers any suggestions for the computer vision AI solutions for developers well if you’re a traditional software developer then I would just appear a software developer.
that doesnt have any sort of you know machine ran background then I would encourage you to go take course ai its nei course for soccer engineers is taught by Jeremy Howard and it is evidence by far the best way to learn AI out there far better than anything else youll see on Coursera Udacity fast buddy on cannot recommend enough you wont get more than you can ever possibly use ever daughter within wet at AI service what is a good price point considering that initiative one may have to use a cloudy I search for an in-house.
for analysis are feasible so almost all your cost associated with AI is going to be data and people so it really comes down to you know kadhi of the people that can either configure the algorithms for you or do you have the data they can use to make predictions and there is this kind of um you know this is this industry idea that oh my god is a.
massive shortage in AI as a massive there’s not a massive sort of atomically this a massive sort of talent for every tech skill right its hard to be good job of the developers but there’s also a lot more Java projects out there there’s more people who are Java developers there are fewer people who know how to do AI and its also a fewer projects so I would encourage you to start and Im going to meetups meeting people look for talent that can help you get your project off the ground we talked about.
price that’s what its gonna fall into the computer resources um no its much cheaper at the margin to build your own deep learning boxes is to run Amazon Cloud but you know its really its its not your major cost I know that you have done work with insurance and financial services nobody its more typically risk-averse what area of the business honey screams dont want to accept that risk and jump into the AI waters are they starting small as something like competitive search for illegal or jumping into the deep and with deep.
learning and under-lighting I unfortunately I cannot answer that question so excellent question I think if you looked around the UM I would my suggestion you just google the question and look at a site called cattle calm and look at the projects that are going on its all over the place you know were working with some of the companies in finance were doing something ratably innovative work that is really going to be you know disruptive and this is incredibly exciting and youll find other situations where theyre trying to.
automate their sales and marketing processes using techniques and tools that have been around for a couple years and so I guess what I would say is for those kind of questions start thinking about AI as a fundamental technology think of it like the way you think electricity or the internet or computers you dont ask how an industry uses the.
internet and because its so ubiquitous and that’s the way a AI is going to be is a fundamental building block that’s going to change everything so almost anything you can imagine using technology for there’s somebody at one of these companies thinking working on it okay I get some answers from some folks here well have questions do you think that unstructured data such as emails and a social media post would have any major effects big businesses and what is the best way to deal with such type of data its an NLP problem so.
natural language processing and you will find tons big samples and how that will have effects so that’s where unstructured data that is text data is an animal pee problem lets appreciate some economy dr Dewalt unfortunately dr the walls my mom actually I fled after my masters degree but thank you what does it create expect that AI would be available in the Middle.
East I assume that its already available in the Middle East I would imagine Ive met people from their region who were researchers are building products and so I would say its its already there see the last couple of questions what are the real application of any industry could you recommend reading materials to study a so again fast that AI is the course that I most recommend and applications and theyll P tons I think I mentioned and.
in the end the presentation and its more in a book but yeah NLP is one of the hottest areas but for the most part any business situation where somebodys generating or reading text is an opportunity for NLP computers are going to be all that force lets see I think I answered this what do you think on structured yeah that’s what I just meant to me that’s the last of our I think that is the last of our our queue many questions and weve got weve got about.
four minutes left Mariah did you want me to okay some more coming in what would you recommend for starts from third world countries where the data is not available as in other parts so I would say that training data is a challenge everywhere even the United States startups dont have data youve got to figure out how to partner with or work with a bigger company to.
get access to proprietary data I wouldnt start looking for if you’re looking for a unique data set for your Legion and I know I lived in Asia for three years a little bit a lot of entrepreneurs and Asia so Ive got some experience in the startup community there I know youve got a lot of challenges but I would just say look around what unique training data sets.
can you get it what is your government have whats being published you know are there datasets there an analogue reform that you can turn into digital form so maybe your opportunity is not actually doing a I and the Train David maybe your opportunity is to create a training video if you live in a region where that doesnt exist Andy okay that’s the last of our questions anybody else that another question okay that was the last three minutes Im gonna make you endure my quiz so were gonna go through this real quickly you.
know what is machine running were some pros and cons of machine learning so machine learning is a different way of building software where you take training data and you give it those examples to an Marlon teach you how to perform a specific task the pros are machine learning is great at abstracting at more and more complex logic it tends to be more robust to changes and as you.
get more complex situations it gets easier to manage the cons are you gotta get all those training data and its a simple problem a lot of time to see if you’re just a write and regular code machine learning is not a magic wand for all your problems another what are your thoughts on the emergence of deep fakes how can we ensure that the AI is youth.
ethically it becomes more prominent Society so I think this is a really good question and fortunately the ethics would be an entire the presentation so Ive seen a deep fakes um you know I guess well see I cant see how this goes but you know weve had if you listen you know in the United States right now people dont seem to believe.
the facts and so I guess or they want to make up whatever facts they have so Im not sure deep fakes we have now I mean its getting easier for people to doing fakes and video but you know weve had Photoshop for decades and well have to come up with ways that are making authenticity if there is a fate was ever being challenged with claim its a fake you can already do a lot of photoshop and unfortunately I think we live in a.
situation now where people are going to believe whatever they want to believe and ethics is a really interesting promise when we think a lot about I dont its more than I can go into in 30 seconds Id say if you’re looking at the real sort of ethical issues when Im AI the biggest one that I see is job loss and then as automation the job loss is coming and we all have to be ready in.
democracies we respond to it okay I think we are just about at our end Bry anything final for the group before we go I would just like to say thank you so much Kevin your time weve really appreciated it and also to everyone that is still listening this webinar will re seminar will be available online if you look under our learn tab and this will be under our how to knee seminars so you will be able to.
find it later dont worry and yeah just thank you so much Kevin its been so wonderful to have you on thanks everybody really appreciate thanks everyone at MIT for organizers you made this very easy to grade shadow.
Leave a Reply