how's it going my friends I am so excited to be back here is this on am I good ok they're giving me thumbs up I'm excited to be here how you doing my friend I'm doing great how are you why don't you introduce yourself my name is Eric Boyd I lead the AI platform team of Microsoft so the first question obviously in the most important that everyone has for you is are the robots taking over and how can we become their friends the robots are totally taking ok ok no I mean that question is such a great one to get I've gotten variants of that and I hear some of these variants by the way please get your questions in because I'm sure I've gotten some really weird questions about AI you know I was at an onyx conference and onyx is the open neural network exchange it's something we work together with Facebook on and Amazon and NVIDIA and invited the whole ecosystem in as a way to exchange you know models from one format to another but in a space it's a file format yeah and after sort of talking about onyx I got a question that was well have you ever had an onyx model you know like become self-aware and I'm just like it's a phone format that's really not how this stuff works so it was yeah there's a lot of misunderstandings around what AI is how it works and you know what people can do with it and so that's a lot of what we need to work through is hey I can do some really incredible things particularly in the areas of vision and speech and those types of areas but you know it's not sentient it's not self learning like if you have an AI model that plays chess really well it's not gonna clean your floors no I've tried and it made a mess and my wife got mad yeah that it made a mess I bet it didn't do it did do anything so here's here's a question for those that are out there they're like because there's a lot of AI they say/i that all over the place for businesses out there that are like ok I I think I need to take advantage of this where are some good opportunities to start yeah that's a great question I spend a lot of time talking to companies about what are their opportunities and what are the things they can do you know the the great thing about AI is it really is touching and transforming each and every part of every business and so even like the feature that we showed today some of the key notes the Intel Akkad if you think about that that's such a small feature in Visual Studio right just reordering the way that the intellisense or options come back but they were able to make that better and and you know by using AI and having a predictive model that knows what am i likely to be using it for they really made Visual Studio better how much better a half percent better but if you can do a half percent across all the different capabilities you can make so much better and so you know I was talking to one company it was a you know a beer manufacturer and I said well how are you using AI and beer manufacturing and his answer was we don't use it at all in the beer manufacturing we use it in everything else in logistics in the supply chain and predicting the inventory that we're gonna need and pre and forecasting the prices we're gonna pay all the other capabilities that they had they were using AI models and those areas of their business and so really it's it's that's what's so interesting to me is just finding each business and how its gonna transform each of their businesses and really challenging companies you need to have the creativity to figure out how are you going to actually make your business better where can a I really work for you and really learning those different areas so what is it good smell humor code smells I don't know where that term came out it was such a weirdo know that term it all his code smells like you're looking at code and you're like this smells like something we could fix it or make it better where are the AI smells so to speak in a business that you're like hey you know maybe you should consider using this program because I consider AI a programming technique that's right you know what our good clinic cases for it if you have something where you know you are you're applying some rules like the sorting algorithm and Intelli code right someone came up with a rule that said let's sort it alphabetically you can almost always find an AI algorithm that's going to predict or do that better fraud detection is a classic place where people have rules if we're get too many requests from this IP address is probably fraud AI is going to do fraud production prediction way better if there are you know financial forecasts I'm gonna look at these three things about you know the the industry or the trends and sort of project what our finances are gonna look like a is gonna do that better insurance industry is getting completely redone if you think about how insurance works today they look at five different factors about your driving history or your car you have or maybe your neighborhood and a I can consider ten thousand factors effectively okay and so those are things where a is gonna do a much better job of predicting and so yeah those are all things that smell like a I would do well in it so tell me about the requirement the data requirement for a I like cuz a lot look I've had people come up to me be like hey just do some AI and I'm just like well can you give me some data why is data such an important element of AI yeah one of the first things I talked to companies about when I meet with them is if you don't have differentiated data you don't have differentiated AI so what is the thing that you have that's special and unique about your business that's gonna make your business and your product better and so you know if a company is not collecting that today and a lot of companies I've talked to companies and they're like oh we throw all that data away and you're like that's gonna be the foundation that you're gonna build your models and make things successful from but you've got to understand what is that differentiation what is the thing that you've got that's unique about your business that you can build a great date asset on and then build models on top of that so before you even start thinking about AI you should probably think about saving your data just making sure that and saving collecting and really looking for places where you have data where you can tell this was a good outcome or this is a bad outcome right if you take insurance this person default on a loan this person didn't default on a loan so I can I use that to predict forecasting financials I knew why I predicted very accurately based on this information I predicted poorly here the more you have those those are the labels that's the stuff that's gonna feed into your model and it's really gonna you know be the way that your AI algorithm is gonna learn and so if you have that then you've got the foundation to do something interesting awesome so you're using the term a lot model yeah and for those that are maybe don't know what a model is how would you describe what is a model yeah a model is it's really just a system of predicting what's gonna happen based on some inputs and so I mean you can think of a really simple model you can think of Scott today Scott Guthrie was talking about I want to predict the price of a car and so a model could be as simple as you know I take the year that the car was and I take the condition of the car and I multiply it by three and that's the price of the car that would be a particularly bad model an AI model is now going to take those same inputs and come up with a price based on you know how a I sort of learned and so there are a lot of different ways you can represent models and think about them but at base that's what a model is it's something that predicts something awesome so we want all of your questions so make sure you get them in and I want to ask them for those that maybe don't have a lot of data and they want to do AI how can Microsoft help so there are a lot of ways that companies can get started there are a lot of use case that depends on what people are trying to do if you look at their speech you know if I want to add speech recognition to any application that I've got how could I do it well I could go and collect a whole bunch of speech people talking and transcribe that and then build a model that predicts that or I could use a card in my service which is gonna do it way better than what they're already doing so we have a wide suite of cognitive services that'll just really help accelerate people you don't need anything to go and use a cognitive service and you can do speech you can do vision you can do language you can do search all of these you know right off the bat you can go and get started with them yeah and it's really easy I have it here on my screen you go to a sure calm front slash cognitive I think it'll go right to it and like an easy one to look at is his vision basically I'm clicking it right here and if you go to for example seen an activity recognition and images you basically just upload your picture yeah it's really simple if you have you know you want to be able to identify the objects in a particular scene just upload it and it'll sort of show you the different things on it then you can go further with that the custom vision service and you can say everyone's favorite example if you watch the TV show Silicon Valley not yet but I will hot dog or no hot dog hot dog no hot dog this is a classic app really funny part of the the movie or the TV show you can build that in about I built it in two hours but that's because I'm not a great programmer anymore I usually happened I got old yeah I stopped programming as the main you know two hours though to build a hot dog non Hunter and it was really very good I did you use cognitive services what I did is I use the custom vision model okay and so I went to Bing and I searched for pictures of hot dogs and I found about 20 of them and I put them in there and I four pictures of things that weren't hot dogs nice and there's your classifier and it really was very straightforward to go and build awesome so here's some questions coming in does Microsoft AI support self training classifiers so I think we do with Auto machine learning so automatic machine learning will do that automatic machine learning is yeah so I guess it depends a little bit what they mean by self training classifier automatic machine learning is gonna do a lot of that what automatic machine learning does is you basically would specify here's the data set that I have and I don't know do I need a support vector machine do I need a neural network to a needle gistic regression I'm not an AI expert please just figure out the right model for me and so automatic machine learning is gonna do that for you and so that's available today you can get that as part of Azure machine learning in the SDK it's one line of code and it'll go and build you a model just based off a data set I have the line of code actually right I'm scrolling it here it is it's it's literally oh no that's the deployment sorry that's this here it is it's the auto machine learning yeah it's actually two lines but because one you define the problem and to run it it's actually one line right actually really nice it's really straightforward really simple to use and we're finding a ton of customers who are using this and it really broadens the scope they're really two ways I see people using it one is you know if you're a data scientist there's a lot of tedium that you go through in trying let's try the support vector machine let's try a learning rate of 0.1 and trying to change all these parameters now you can just sort of say just give me a model that's pretty good you know an automatic machine learning will just go and do that the other class of users are people who don't even know what they would need to do and they can you know if I'm a developer but I don't know how to do machine learning I can use automatic machine learning and get a model just from two lines of code so let's see if I understand so basically you have a bunch yeah III obviously usually I think of Excel because it's the canonical I say Excel and everyone's like oh I just pictured a row like a bunch of square of data and you say here are the columns I want to use to learn and then here's the column I want to learn you feed it in and then it does it and it's funny you say excel because in power bi in and we announce this as in preview that it'll be integrated as a wizard where you can walk through the experience doing exactly that and so create a model just from saying this is the column I want to predict and here's the rest of the data that I've got and that's pretty amazing I mean if you look at the code right here hopefully you can go to the screen you can see that there is this data script that's basically says get data and in there you define and you can see my get data script right over let me go here real quick get data script right here it's basically saying here's the X and here's the Y you want to learn you give it that and then it says I'm gonna run a bunch of model that's right it goes through it tries all the different models it's really great it's a it's a little meta it is a machine learn model that predicts which machine learned model is gonna work best on this set of data and the more you use it the more you use it on your sets of data it actually learns right from the history of how it's gone and predicted and it gets better because it has a great answer of like well how well did it predict against the test set and so the more you use automatic machine learning the better it's gonna work for your data the better is gonna work for you which is really pretty neat and the cool thing about this and this and this is the meta bit like it's using a recommendation algorithm so basically like your shopping cart when you go shopping online and it recommends other products we have that AI when you run a model and it says oh it did okay it starts to recommend others and it runs them it's pretty cool it's exactly it's the same idea so the exact same implementation was the same ideas that yeah really cool really simple to use and I think honestly going forward I think more and more stuff is gonna start fitting into that automatic machine learning paradigm I think you're gonna start seeing image classifiers work that way you know I all the types of things that you're doing today more and more of it's gonna fit into automatic machine learning and that's that's really our goal is how do we simplify this stuff because there's so many people now who the demand to have a machine learn model the number of people who wanted the demand freya is really high but there aren't enough people who can go and develop it and so how do we make it simple so that any developer can go and build and have an AI model you have developers today if you think about I use the analogy of a hash table developers use hash tables without even thinking about right they probably learned about it in college and they studied what's a good hash function and things like that no one's ever actually implemented that they just use a map and they just call it right how do we get that you know machine learning from the days where it is today we kind of need to be an expert in linear algebra and calculus and you know back propagation to know this is a space here's my data push a button and go do it it's coming and we're gonna get there but you know what it's still okay to learn linear algebra or calculus because it is dear to my heart because I'm a nerd I'm not wearing my glasses but if I was you would see love calculator oh and and this is what you're seeing a lot of people doing everyone who's a developer out there today is learning more and more about how to be a data scientist and it much like the hash table understanding the fundamentals is really going to be important to doing it effectively awesome so it looks like Tony from Brazil is watching we want to say hello to you my friend I have an idea in mind to build a face recognition for my business customer comes profile loads for receptionist can azure help me with this that's a great question so we have the Congress service the face recognition API and so you can absolutely go and do that and have it recognize your customers and and suggest a profile of it I think it's you know it's really great idea right a customer walks in and you say oh they typically order you know this particular type of cup of coffee or something you can start getting it ready as soon as they walk in and so for your regulars what a great experience right I'm coming back to that restaurant every time because they're gonna be ready that much faster they see me coming and that's amazing like just with computer vision it will tell you the box around the face right and then you can use in China together with with custom vision with faces of your customers and so you get the face box you take the box out and you give it to custom vision and now it'll tell you that's right you can build up the profile in the history over time and so I think it's a really interesting use for how face recognition they're actually already are coffee shops and China that do this really so absolutely you you walk in and they they recognize you and they get your drink already so I think this is gonna be coming more and more awesome so follow-on question but do you have an auto classifier where the data does not have predefined data so the auto classifier would group similar things I mean that's a really good question I'd like look I'm an engineer and most people don't know this but I'm like an AI person yes so like right away when you say group things I'm thinking like k-means or hierarchical clustering or Gaussian mixture models so there's already ways to do that yeah but we have the compute to do any type of machine learning I'll tell us a little bit about Auto a machine Azure machine learning service and what that is sure so you know you talk about you need access to the compute to go and train your different models that's one of the you know why is AI really taken off it's been three factors first is the amount of data that's available the advent of big data and product you know products like spark and an azure data bricks and now you can use those to really manage your large data sets the second you alluded to is compute and it comes in in two different ways one is it comes through GPUs which have really accelerated the amount of parallel computation you can do and things particularly matrix multiplication as well as the cloud where these GPUs are really expensive and if I need twenty or a hundred of them I don't wanna have to go buy them because I probably only need them for a few hours and so the cloud ability to go and use all that compute you know for the time that I need and only pay for what I use has really transformed it and then the third thing is the new algorithms coming up with you know deep learning and convolutional nets and things like that has opened up a whole host of applications and so with Azure machine learning what we're trying to do is bring that together and make it really easy for you to consume and so you know we do it through a Python SDK so you can be in an azure notebook or a jupiter notebook or in vs code or in pycharm or whatever you like to use you're right there in your notebook and you can get access to start training locally change one line of code you know now i'm accessing a whole host of resources on the cloud and you know managing the data that's stored in anyway you can store it on Azure data at everything from Azure data Lake to you know cloud DB everything is all out there cosmos DB awesome so yeah it's all really simple and integrated so we started with if you want to just get started quickly you should use cognitive services yeah if you want to customize a little bit we have some custom versions of cognitive services we mentioned custom vision there's also others you can I think you can upload your own acoustic and language models for speeds that's right you can use Lewis to start to tag your own data and get so there's ways easy ways to enter but let's talk now to those that maybe are more advanced that are starting to use like scikit-learn or pi torch or tensorflow can you tell us what the current development process is like and how Azure machine learning trimmers can help sure you it's interesting the progression you walked with we think about it as cognitive services are you know Microsoft's model Microsoft's data the custom services are Microsoft's model with the customers data and then when I want to get to a custom model well then it's the customers model and the customers data and so you know how do you do that if I'm building a model you know frequently what I'll do is I will come up with the set of data that I have that I think has the best features in it and I'll come up with the algorithm that I think I'm going to use and I'm gonna Rigaud and train it and then we call that an experiment and I go and run and I compare that experiment against the the history that I've got the test data that I've got and I see how well I performed and what you find people doing is this iterative loop over and over where they keep changing maybe I need some more features or made this feature is gonna make it predict you know in my car example maybe the year is not very interesting but the location the zip code is or something like that what features can I add in and so as your machine learning will keep track of all the experiments that you've run so that you can see which one actually perform the best without and you know basically keeps the history of what did you do different each time that you had it and then when I have a model and I'm ready to sort of deploy it then the you know Azure machine learning will put it in the model registry and so now I can keep track of all the models that I've got and as I deploy them I can understand where each model is and its life cycle is it deployed in which different systems and so it can manage that deployment and management of them much much simpler there a couple of other capabilities that I think are pretty interesting I talked about sort of the the the lifecycle of I'm going through all these different experiments a lot of times what I want to do is what's called hyper parameter tuning right each of these models there are probably a dozen different parameters the learning rate the number of nodes in each level all sorts of different things and so hyper parameter tuning is a way of helping you select the best ones without having to sort of painstakingly iterate through them all yourself and so you put all that together and it dramatically simplifies the machine learning developer the data scientists in getting their models develop they're become much more productive we talked to people who you know have spent weeks training a model and rather like we can get this done in a day I can do this much much faster and the important thing to note is that like look generally when I'm running these things I usually run it on my machine and then my machine is tied up and then I try to get someone else and I'm using PI torch 0.4 0.1 9 7 & 8 you know and Sally's using PI torch 1 preview because she's all way more advanced than I am and it's hard to get all these things to match up how does a skirmish e-learning help people that work in teams and data science teams what is it about Azure machine learning that will help people work together yeah I mean there are a couple things you can look at you know one is you mentioned like my machines all tied up that's the beauty of the cloud right is I can now use extra computation on the cloud I can put it in a data science VM I can put it on the training modules a whole host of GPUs and have sort of complete access to my machine the other is really the reproducibility I can take sort of each model and I can sort of run it the exact same way that you ran it and so yeah you mentioned the different versions and all the different things that I'd need to do you know makes it much easier to sort of say hey we're all sort of working on the same thing and share the code and and really stay aligned with it and that's something that we've learned internally you know we have on our internal services thousands of developers working on the same model and Bing and how do you do that effectively how do you have a thousand people try and make improvements to the single model you have to have all this infrastructure and so as we've learned from Bing all that infrastructure that we built is through our machine learning and that's what we're deploying in as machine learning all the things we've learned in making that product much better awesome well keep your questions coming obviously any AI questions that you might have we want to get those in so there was announcements regarding AI today can you tell us about those sure it's a couple announcements that we made first and foremost our machine learning is generally available I'm really excited about this we've had a whole bunch of customers try it and preview and their feedback has been really really great they think the direction that we're going with it has been fantastic we feel like the quality of the service the quality of our documentation understanding is all great and so we're happy to announce that it's generally available product and you know with that all the things that come with that automated automatic machine learning which we've talked about is generally available the hyper parameter tuning the experimentation capabilities the model management the ability to deploy all of that service generally available and something that people can go and take and use which is really exciting another thing that we announced today is the onyx runtime so I talked a little bit about onyx at the start onyx is this file format for exchanging models and you know for really making it simpler for hardware manufacturers to to optimize them one of the challenges that hardware manufacturers talk us or talk to us about is they say look I've got tensorflow I've got PI torch I've got chain ER I've got paddle paddle I've got cafe I've got all these different frameworks and people expect me to optimize each and every one of them how do I do that simply and onyx says hey you can just sort of make this in onyx model they can all convert into onyx and now you can optimize one of them what the onyx runtime is is this is the same runtime we've been using in Windows and Windows machine learning it's now available open source it runs on Linux it will run on Windows and it runs dramatically faster than sort of the native implementations and so we've seen internally you know virtually every model has run faster some have the average is probably around 2% two times faster but some have been as many as like seven or eight times faster and so this is available it's open sourced and people can go and get it I'm really excited about that as well another thing that we announce that there's sort of major areas around cognitive services we are the only company that does containerize Congress surprises people want to be able to run AI models and they don't want to just consume it in the cloud sometimes they have latency requirements or they have intermittent connectivity for some machine that they're trying to run it in and so they want to be able to take the model and run it on Prem or on the edge right and so we announced containerized models and today we announced that the language understanding model is also available in a container amazing Lewis Lewis Lewis is now available in containers so that's amazing I was recently like I was recently in in New Zealand and one of the hospitals there they wanted to use cognitive services to do OCR yeah but they couldn't do it because they could not but the records could not be upload to the cloud this really enables them to start to do that work that's exactly one of the use cases we see is you know regulations prevent them from moving the data and so being able to bring me the cognitive service directly to where they actually have the data opens up all kinds of doors that previously they couldn't do before and so we have had a ton of positive feedback from customers on it people are really excited about this so I'm excited to see that going out it's basically the first ever lift and shift down that I've ever seen yeah it's interesting all the you know everyone is trying to lift and shift up to the cloud and and look that's an important trend because the cloud offers a lot of advantages but as a company we've been very committed to the hybrid and making sure we work with people where they need to be and often that means hey some of the things we've done in the cloud need to happen on the edge there are real legitimate requirements a little why that needs to be done and so we want to make sure that that's support like just basically and here's a really silly use case that I thought of if you run a parking lot and you want to have an unattended parking lot and still charged people you can have a local model taking picture as cars drive in with OCR getting license plates that's right all sorts of examples like that where you know we hear of manufacture of companies that have manufacturing plants that you know their connectivity comes and goes right they have vehicles that they want to drive out into the field that they might not have any connectivity at all and so to be able to still run models locally these did a lot of great examples with drones they want to fly drones along power lines and sort of see hey are there defects and the power lines and things like that things they need to go repair you could have someone drive for thousands of miles you could fly a drone and have it take pictures say this is where you need to go to that's awesome and so yeah it works great so here's a couple questions well ml dotnet expand to training framework so I don't have to learn Python numpy tensorflow PI torch and I can rather stay in c-sharp or I feel at home so I'm all that nut has a whole host of algorithms where you can it you can train models directly in c-sharp in ml dotnet and well that net again grew out of internal technology that we had were a bunch of internal developers use C sharp and wanted to have a good framework for developing their models in c-sharp and so we've wrapped it up and made it ml dot and that is something that now you can go in and train models in c-sharp you know we have a ton of developers in c-sharp that are really not well served by the ml community because the ml community says everything has to be in Python and so being able to have ml net as a way that you can now stay in c-sharp is really pretty interesting for people and you know there's still a ton of activity happening on Python so I wouldn't necessarily discourage you from learning Python because that's where there's a lot of value being created but absolutely I'm Ella and that's gonna let you stay in c-sharp awesome next question ai is more than just image recognition what about searching for text and dynamic images kernels your engine is not accurate and cost per transaction forces us to use on-premise solution interesting so I'm not really sure what they mean about not dangerous or dynamic images yeah I mean maybe in videos or something I mean so the the OCR solution that we have I get benchmarks on all of our under services on a weekly basis we're beating everybody else with OCR so I believe we have the best OCR solution that's out there it really performs quite well and you know we have I talked to our researchers they are very proud of the techniques that they've used some of those some of the tricks that they've done in their models to make it really perform so well you know additionally with images you need to sort of talk about finding different things within the image I mean I love the example today that you had the NBA example showing really the face detection working on all sorts of small areas there's also logo detection right and being able to find the logos and all sorts of different places in it and one of the places where we really pull all this together is a knowledge mining right so you take exactly what you were doing and now you've got your extracting all this information from across an image and some of it is OCR some of it is looking for text in an image some of it is face recognition on the image some of it is sentiment analysis on the image or on the text or whatever it is I'm building an index that's now searchable and brings all the you know the named entities you can sort of think of whether they're people or places and what's the relationships between them you can bring all that together and so knowledge mining makes that really powerful now and I will say for for you that that submitted this question like if it's not working for you can you email me and I'll look at it because I've seen our OCR engine get like words behind chain-link fences that say stop that are crooked and give you the right bounding box it's ridiculous I've even seen it there's you love the blacked-out text where they used a magic marker and you can still kind of see the text interact of it and our OCR is telling you what the text is underneath of it which is really very cool it's amazing okay so tell us a little bit about some of the exciting things that you see some of the customers doing because one of the problems with this technology is people have a hard time like saying they once they get it they're like oh okay I can see how this works yeah but they have a hard time seeing where it might fit in there in their business what are some exciting things you're seeing customers you know there's been a lot of interesting examples you know one of the things I may be more fun ones is you know we've talked to shell and so shell has gas stations of course all around the world and so they put cameras in there gas stations and what they really want to detect is fire which is a big problem at a gas station but what they also are detecting are is the person who's filling up with gas is he's smoking and if they are they want to alert the attendant say hey go yell at that person it's really dangerous to smoke at a gas station and so lots of applications like that we see a ton of image classification applications companies you know Jabil is looking they manufacture chip boards and so the last step they wanted to sort of take an image of it and see did we miss any of the solder is in it is there anything any defects we can see with this and so just improving the quality that they have there we see a ton of Bott solutions people are looking for BOTS you know often in customer support can I deflect you know twenty percent of my customer support costs and save myself twenty million dollars while giving my customers a better experience we see virtual assistants where people want you know their branded experience their voice you know sort of interacting with a customer in a virtual assistant way so we see a lot of solutions there so yeah just a ton of applications across the board awesome so we have a question coming in how could I use AI to help with accounting such as invoice coding so I'm not an accountant and I'm not entirely sure what invoice coding is but I'll guess um you know presumably it sounds like a classification problem right I have an invoice come in and I need to decide which budget do I charge it to or which type of invoice it should be or anything like that there are a number of different ways you can do it depending on sort of higher invoice comes and one is you can do image classification if you have an invoice that sort of looks like image one or an image and something else that looks different you can do it that way the other is to build sort of text classification and sort of understand what type of system this is but that's the standard classification problem and that's the exact example of what I'm talking about of each business and each industry needs to find the ways that things are changing for them and how they can really use it to make their business better where you know if you can speed up your your classification on your accounting invoices and have better accuracy with that you just get so much acceleration from that as a business and so really finding those ways that haven't really getting the industry to have the creativity to understand one of the places they should be using this technology that's gonna be one of the most exciting things over the next few years I found earlier Mike early in my career as a programmer anytime someone wanted me to optimize something I would look for tasks that someone repeated exactly the same way off that's right now I feel like when I'm doing AI my particular smell is if someone is like altering code like point seven to 0.5 or adding extra if statements and then deploying yeah like for me when you're doing those kinds of tweaks where you feel like you need to take a shower afterwards as a coder that's a good place to start thinking about using a on they're using rules to make a prediction that's what you're sort of describing should it be 0.5 is the threshold should it be point 3 isn't threshold and you know this is where I say you saw this in fraud a lot and I get a thousand requests from the same IP is probably fraudulent and you know I remember the early days when I was working on fraud you know AOL dial-up modems were a huge problem that tells me how old long I've been working on fraud because they all came from the same ip address and so you shut down all of AOL from their one IP address and they and I are models gonna be much better to learn that hey they're different patterns that I should be looking for and so instead of having these broad course rules I'll have these you know I can really pull in a thousand different features in and build a model around that let's talk a little bit about sort of the elephant in the room and this is important because people are looking at this and they're like hey how much is this gonna cost if I want to use if I want to build my own machine learning model and Azure machine learning service is this gonna cost me a lot so the beauty of using the cloud is that you use only what you consume right and so the cost for GPUs if you were to go out and buy a whole host of them they're tremendously expensive and so the rates that you can use when you're using the training service you know is measured in you know the tens of cents per hour and things like that so dramatically cheaper you know building models yeah there's cost associated with it but it's it's not exorbitant this is something and the benefits that you get from it on the other side I don't think I've seen anyone come back and say hey we're not gonna build this model because the the training costs are too high and here's the thing like I I've laid down an azure machine learning service workspace and we basically lay down for things we lay down a pin sites which is nothing yeah we lay down storage which is empty yeah right we lay down a CR which there's a free version yeah right and then we lay down so app insides Azure storage ACR and there's there's one more thing that I'm forgetting happens science as your storage ACR and there's one more that'll come to me and compute back oh yeah the computer yeah the compute stuff and like you and even the compute isn't laid until you specifically specifically asked for it and so you can also create a compute environment that has zero to n nodes yeah and it won't even run yeah I mean that's you know I honestly I don't know how everyone doesn't develop on the cloud these days the the economies of scale that you get from being in a cloud are just they just give you such a better efficiency from that it's pretty amazing okay so we've got about three or four minutes left where can people go to find out a little bit more about this AI stuff and how can they like get start like if you're a programmer and and because you already know all about all this AI stuff what would you suggest someone go do right now I mean the things that the easiest way to really get started is to go and look at the azure notebooks and start with they're the azure notebooks will walk you through how to use Azure machine learning service notebooks are such a great environment for learning in because there's a description of what you're trying to do right in line with the cell that executes the code you can change the code and continue to execute it and make it just go right there in line that's the thing that I would recommend everyone go and do it'll both you know depending on sort of what you need to learn if you need to learn how the services work they're great document they're great notebooks sort of walking you through how to use the service if you need to learn how to build AI models there are notebooks that'll show you hey here's some of the simple iya models do I want to do like you know the classic amnesty I can type draw number and sort of have it recognize you can those are really straightforward ways to go and do that I just remembered it's key vault is the last one so there's four things we lay down storage app insights key bald and Azure as er a CR right which is as a container registry okay so I figured I'd show just a little bit of its what this looks like because a lot of you are probably wondering what this looks like and like it's really simple and the cool thing is that the part that maybe we haven't seen in Francesca's showed that a little bit during Scott's demo is we actually have these amazing integrations individuals to your code where I can go in and submit experiments by right-clicking right or viewer experiments or attaching right and and that's that the coolest part right and the other thing that I really liked about this that I've used so far there's a team of like four or five of us that work on these models and all of us can see what all of us are doing yeah like here's an example that you talked about hyper parameter tuning yeah it ran 20 experiments and notice that here there's like this green line it actually stopped running it because it's like yeah this one didn't work what a type of rider is not working and that's cool because initially like usually I'm running a for loop and spanning over hyper parameters like the learning rate or momentum or whatever here it's a smarter for loop because if it's within and I use the bandit method if it's got 20 percent it's gonna kill it now and that's a great example of you know usually you'd have to sort of be sitting there and watching and tracking having it basically printf hey this is what's the learning rate how it's converging you know high parameter tuning is just gonna do that for you and so it's it's really saves a ton of time for you and the other thing is and this is the part that's really cool here's all the computer environment you see on my computer we have like some batch AI we have some kubernetes services in our compute and it's basically we just submit and forget and then we can see all the output it shows us all the output in notebooks it's amazing now here's the the part that we were talking about models to me are like the executable part of a on yeah that's right it's like it's like when you compile your code to get this assembly I feel like that's what a model is - it was a great analogy you put something in and something comes out and as a programmer it's something that we want a version yeah and in here you can see we are versioning all of the models that's right and then we can marry models with scoring files to create images and then we can do deployments and so for example you can see the simple m nest service right here I wrote it like a cheesy little app here where I submit like I can draw a number this is the cheesiest version the cheesiest thing but you can see it returns things from me directly from the service and I'm gonna make like weird numbers and you can see that oh it really thinks it's a 4 but if I start to do crazy things like this you can see that it's gonna start to be sure about other things right and it's it's pretty cool that you're able to go from idea to submit a job to save a model to create an image deployment all within the same environment yeah I think it's really important that the software development lifecycle for models is different than for software so it is and so understanding what are the tools that I need and how do I really use that more effectively those are the things that are gonna make you productive and successful in building your models and deploying them if you if you try and do it you know sort of the standard software way you're gonna have a hard time figuring out how it all fits together awesome well anything else to finish up with my friend I'm really excited that Azure machine learning is GA I think there's gonna be a ton of amazing uses for it and I'm excited to see what people will do with it well thank you so much for being with us Eric all right well the show is not over we have my amazing colleague Brian Benz just over there with Beth to talk all things java let's go to that
No comments:
Post a Comment