AI Hub: The One Place for Everything AI (Cloud Next ‘19 UK)
How's. Everyone doing. Welcome. Welcome I hope you've been enjoying day one of cloud. Next so far my, name is Nate Keating I'm a product manager on Google clouds, AI platform, focusing, on AI hub which I'm really excited to talk to you about today. Before. I get in just, a reminder that you can fill out a survey for any session, that, you go, to or attend today the, surveys will open in the Google cloud next UK app about. 20 minutes in it's. Hugely important for us right this is this is how we get to understand what content. You value how, we can iterate and improve each. Of our cloud, next. Conferences. Every year, and. How I can actually help improve my, presentations, when I'm talking to more, customers are interested, attendees, so, I thank, you in advance I'll read all of the comments that come out of the session as soon they're available I promise. Before. I get in I kind of want to pull the audience love it do a little straw poll how, many folks would say there what, I call an ml builder, like a research scientist, data scientist, machine. Learning engineer you. Actually work on creating some part of the of the machine learning model getting into production okay cool how. Many would say they. Are, managers. Of either, individuals, teams organizations. That work on machine learning. Awesome. And, how many are, just neither of those and just like love to come, to these sessions learn about AI weather, friends do AI cool we love you too on the cloud platform and we hope to see you around if you're if you're interested. Last. Questions how many are currently using or have checked out AI hub before. Awesome. So a lot of new folks at hub is to be really great it's, a great introduction it's a product we announced in beta just a little bit earlier this year so, excited to share with you a little bit more. Brief. Walk through the agenda I'll, go through an intro kind. Of scope out where. We sit in the cloud ai world, the cloud platform, some. Of the pain points that we saw and we're hearing from all of our customers, that, led us to the vision of AI hub which is in part inspired from talking to folks like you and in, other parts inspired, by a lot of the great work that goo internally. Across, the hundreds if not thousands, of people who need to collaborate on production, ministry machine learning that we have going on across. All of our business units and, then. I will pass the, baton off to my colleague Mariana, who, will walk through some simple. Demos to give you kind of like a sense of how you would use the product and bring it to life and I'm, super excited to bring onstage Atos to, talk about how they're beginning, to use a hub to solve some of their machine learning asset management and collaboration challenges. That they have today and. Then I'll wrap up with you. Know kind of like the next steps how to get started and get engaged with us going forward, but. Let's jump in let. Me start with the mission statement you may have seen already cloudy. Eyes mission empowering, every Enterprise to transform, their business with, AI and the, key here is that and. I'm sure you all know this we've, moved beyond infancy for machine learning we're in the era of what we sometimes call deployed. AI where, teams are not just like experimenting to see if they can get something in production, that works but. They actually have models in production already are, figuring, out how do how do I actually improve. These build, continuous, delivery pipelines. And collaborate. More effectively, taking.
The Learnings they have from certain use cases to new data sets and new fields. So. Google we've pioneered a ton of breakthroughs. In machine learning or AI space, but. In cloud AI that. While, we well that those breakthroughs are like a big underpinning, of a lot of our products, what. Really drives our product, roadmaps. And like our product thinking is not just what are the great breakthroughs, but what's really going to drive business value, for, all, of our customers so, we take the best of Google's. Research and all the ways we do things internally and marry, that with all of our customer conversations and, the things we're hearing are the challenges that we're having that maybe even Google isn't isn't running, into across. Your industries. And. Our, businesses broke down to three areas on the, left-hand side we have our building blocks which you can think of just as like. Pre-trained, machine, learning models exposed. Through an API. Example. Would be the cloud vision API kind, of a classic example send. It in an image comes, back with a bunch of classifications. The. Key here is you don't need to bring your own ml expertise and you don't need to bring your own training data you can leverage the power of some AI, without any of that. Already. On the. Right-hand side our solutions, these, are larger, offerings, that seek to take an, entire, like, horizontal or vertical business. Challenge, and provide, an AI powered, solution, directly, there to your enterprise but. Today we're going to focus on this middle section which, is a cloud AI platform, of which a hub is a major part and. A AI platform. Unlike, the building blocks is a place where you bring your own machine learning expertise, and talent, and business context, you, bring your own training, data and what, you're accessing is a huge toolset and a bunch of infrastructure, that, we provide to help you be more effective and more efficient, in building, machine learning models repeatedly. And, reuse ibly. And. So what's the main pain point why do why does the cloud platform exist, pretty. Simply it's really hard to do production machine learning right now it takes way too long most, of the customers I talk to will, say that it takes over six months and sometimes up to if not over a year to get a brand new model into, production, let, alone actually, iterate on that model and improve it going forward this. Image may be familiar to many of you comes, from some of some canonical papers, that the tensorflow, extended, team put out about machine, learning so why is it so hard and, primarily it's because very little of what you do in production machine learning is actually, algorithmic, decisions, or training.
A Model there's a whole host of things you have to think, about and take care of both, in processes, and technology, so you have to ensure that your. Training, data and your, inference. Data is similar. Enough that you're not going to get training serving skew you have to handle auto scaling a be testing, detecting, bias which, is becoming more and more an. Important part of what we do here and part of our explain ability offering that you might have heard. About during our our keynote so. This is just a few of the challenges that the cloud platform tries, to solve and I think the analogy, that really helps here is we, see ourselves in, some ways as like the assembly line for machine learning it's. Not about like, the, innovation, of building a car like we kind of have that not. Entirely solved, but well but solved well enough what, we need to do is build an assembly line that lets us build cars really. Quickly and roll, out new models and new versions of cars really, really quickly with, the way in like 72. Million cars get put out every year, I kind, of think about a world where, across. All the enterprises, were building and putting 72, million new models in production, at some point in the near future that's, the world that the cloud platform is really trying to move towards. And. Of. Course there's a bunch of different stakeholders across, this entire machine. Learning workflow right and we tried it to provide, tooling, and, collaboration. Ability, across, all these different personas, and stakeholders, but today data science is still largely, an individual, effort maybe, not as bad as it was five or ten years ago but, still today most folks I talked to have, an objective have. A data set and kind of go off on their own in like hammer away at it for an extended period of time a few Sprint's or months, then. All that experimentation work, that they're doing they. May have learned a lot and certainly they may have a better model because of it but no one else has learned anything from this and so, if you are interested in figuring out why that model is better than the one you did or why that model performed, poorly, you're.
Basically Asking people and. None of that organizational, knowledge is actually being shared or referenced. Unless your drink's extremely. Detailed documentation, I don't, know if this resonates with anyone in the room but I hear, like. Like, data, signs themselves say they don't even know what why they did things six months ago and so it feels like they, don't they don't even retaining, the knowledge or the experience, of their, experiment, experiments, and experimentation, processes, through. Time and, so, this is the core problem that that we are starting to tackle, and, so you some of these again may resonate with you here is that there. Are tons of knowledge silos, and redundancy, in this world I. Like. To think of many of our product theses this way which is as, organizations. Start. To hire way more machine learning experts, and data scientists, growing their teams and as, we start to offer tool sets that allow each data scientists, to scale up how many experiments, they're running and how many new models they're building very. Quickly we're talking about two three orders of magnitude more, artifacts, being, generated, across all. Of your machine learning processes, right so we're. Going to need better systems, to handle and host and compare. And document. All of the machine learning artifacts, that are being generated across. The workflow not even just models. And, so how do i how, do i improve. My speed of innovation by not starting from scratch every, single time how, do i get a view of all of my organization model in models, in one place may be filtered. By which data set they were using a rich features they're using how do i trace the lineage of each of those models back in time and say like okay what would the what's the exact code that generated this model what, dataset was it run on give me an exact snapshot, of that dataset so I can try and reproduce these results and iterate from there. How. Do I eliminate redundant work if I've got a bunch of, people. In in Europe and us, working on similar datasets trying, to solve similar problems for, their customers, how, do I prevent them from basically, trying the, same exact things over and over again without realizing, that they're they're both working on similar problems, and. This is where a hub starts to come in so a hub, is intended. To be the one place where, an organization can manage all their machine learning artifacts, and therefore. Drive collaboration. And, we do this primarily through three user, journeys, first. Is discovery. Today. Again, if you're storing most of your machine learning work on your local or in the side of Yemen the cloud or maybe, even using a git, based source, repository, it's, still not searched first it's, not intended to put in front of people who, don't already know that that model or that notebook exists, and so we try and build here as an index across all of your work all of your machine learning components, so that people, can search for relevant, artifacts. Whether they be datasets features, notebooks. Models or pipelines and. Make them easy for you to for them to find figure. Out what's going on there and maybe take and iterate on themselves. The. Second is usage right this isn't just an index right I shouldn't just be able to find something learn, that it exists, and then still have to go ping someone and talk to them instead of meetings I should be able to get a lot of context, from this as well so if I search for a model and I find it in AI hub I should have detailed. Logs I should be able to open up a tensor board about this about for, this model I should be able to compare, it with some other models that I have as well and, then most importantly all, of these should be kept like in warm storage, so I can quickly deploy and run and test any of these pipelines, or models or notebooks, in the, future, and. Last but not least is sharing right look we don't have collaboration, unless we have the ability to share things but. This also needs to take an. Enterprise. Focus, view so, while I do want, to, expand, the circle of people who have access to my work and can learn from it and improve their own work I also. Need to make sure that, only.
The People who should have access to these do, and so. Enterprise-grade. Access. Controls and permissions, as well as. Our, standard across GCP. So. Where's the hub live in like the standard GC PML stack for those folks who are already doing their, machine let me work on GCP of, course the core of any cloud, offering, right if you're starting at the bottom here is infrastructure. Compute. And networking. Right like this is the the standard thing that you. Get when you come to the cloud on, top of these though we've already provided. A bunch of great managed services, that. It make the backbone, of the cloud platform and I won't go through all of them but just an example right the AI platform training, and prediction services. Let. You run distributed, training jobs. On. Google cloud or do. Online inference, without any hassle, and just basically, hand over the reins, to Google clouds infrastructure, autoscale, as you need do, training jobs on an insane number, of servers. And do they just spin up and spin down whenever whenever, you need them. But. Most of the scientists probably spend a lot of their time in that like tooling, and SDK layer like inside, notebooks, writing code and IDs using. Some of our SDKs, to. Orchestrate, pipelines, or. Do, some explain, ability work and the. Idea for a hub is that each, of these lower layers are, going to continue to generate more and more artifacts, that, today. Again, just like get, put into a bucket or dropped. Into your local and you, lose context, of why when, how they were created that, a hub is the place where these maintain. Their context, and can, be surfaced, going, forward I think. At this point it probably is just easier, to give you a quick view of what a hub is so, if you can switch over the demo machine I'll just give you like a quick overview, and then pass the past them cough. Actually. You. Switch and Mariana, it'd be great if you came up. So, what you're looking at here is the, home page for a hub once you've logged in and. There's a lot to go over here so I'll try and just really briefly point, out some. Key updates before passing it off so. The first is that this top section looks very G sweet like and that's kind of intentional, right is you've got some quick access to popular assets, that, are uploaded and shared inside, your organization so we have pipe like a training pipeline here we have a couple of notebooks that might be valuable, below. That we have new assets, that were recently shared with you just again for quick quick access, below. The fold though we, actually have a separate section of a hub where, Google has open sourced and provided hundreds, of machine learning assets, that might be relevant to you so, the first section here is production-ready, assets. And these can be some, standard VM images that you might want to use some. Pipelines. That take data out of bigquery and do some pre-processing, or. Even full end-to-end pipelines. Another. Row, here in their shelf for, Kaggle which is now part of google clouds ai platform, for those not familiar Kaggle, is the world's largest data science community with millions and millions of of active, members and. We've actually gone through and curated, the, top Kaggle kernels that we think are valuable for enterprise, and are in the process of doing that on a regular and frequent basis, we're updating this curated, section of machine learning assets, which, made us be relevant for educational, purposes but may also you may, also find valuable. For for building, on your enterprise, use cases and, last, but not least what, would we be if we didn't provide quick access to. All the cutting-edge research that Google, and deep mind and our cloudy eye research, teams are, putting out there I, think. Another thing to point out here across our public, catalog is a few other asset, types so besides machine, learning pipelines, notebooks. Some, of our api's and services, we also have tensorflow modules, so, for folks who aren't familiar this, is like a check pointed, tensorflow, model, that, you can take and do transfer learning on so for example you don't have enough data to do a full to. Build. A model. On your use case you can take a check pointed model from. One. Of these tensorflow modules, trained. Is the last few layers for your specific use case and run, this going forward. I already, mentioned me at VM images a place, for trained models and these would be like the model binary files a like, a TF save model and. Last but not least we have technical guides in the public catalogue which, are kind of like these long, 20. Plus page detailed. Guides and how to set up a more, complex machine learning system, with, an attached, source, code repository. For you to actually just run and get it get it up and running. Couple. Other things to quickly point out just from the homepage any, of the assets here you can actually star.
Which. Is a great way to bookmark them and then, you can see your starred assets just on the from a quick access here on the left hand side so if there is something you find super valuable and want to come back to you can just quickly grab it right there, you. Can also search so let's say I want to look at assets. That use Kerris and I immediately pop up a bunch of notebooks here all, using, Kerris, 211. Of them here which. May be valuable for any of your use cases or for learning more about how to do deep. Learning in, a very fairly simple manner using the Charis API. Last. But not least you get quick access to all the assets you've created in AI, hub right. Here so into the my assets, page you, can see a list of and this is just the few that we have in our demo environment here assets, that you've uploaded with, quick easy sharing, icons, and then icons. To show which have already been shared and which, version, number you're looking at but. Again to any more depth about how to actually bring, your assets into the, hub ecosystem, I'll pass it off to Mariana, to go through some more detailed, demos. Thank, you. I'm Mariana kina Garcia I'm a software engineer for, AI hub and I'm, very excited to show you what, we've been working on for the past few months so, let's. Jump into it let's. Say I'm a data scientist, and I just built no. Book which, I want to share with some people on my team. AI hub allows you to share the source file the, metadata. And version, information for, your asset, in just, a few clicks and. So. For example right now I have a notebook on my, laptop which, builds. A binary classification, model. And. I'll. Run you through that notebook in a bit but first I want to show you how to upload and use it in the AI hub. So. I click. On this top left button. Select. What, I want to upload a pipeline. Model. Or a notebook I'm, gonna select the ladder and. First. I'm gonna enter some, metadata, for, my, notebook let's. Call this. House. Pricing. Classification. I'm. Going to set the input data to tabular. Now. I'm going to enter some optional. Additional. Metadata that will just make my asset easier, to discover, in AI, hub for, anyone, who has access to it so, I'm gonna specify that this takes me and to end in the machine learning workflow, and. That. It uses the. Classification. Technique. Gonna. Type, brief. Summary. Type. Very fast. Click. Next, in in this step I can actually upload. The. Ipython. Notebook, file. The. Last step I preview, how it will look once I actually submit, this as you, can see we render, the notebook and the metadata I, previously, entered, and. Once, I hit submit I'll. Get, a confirmation, that. My asset. Has been published. If. The Wi-Fi allows and. Just. Like that my notebook is now an asset, in AI, hope. Now. What what can I do with this well. As of, now it's only visible to me so to the publisher, but, once on the details page I can just click on share and share. It with whoever I want so, I can either give everyone, in my organization, viewer permissions, or I. Can specify a user through. Their email and give them viewer or editor, permissions. If I want them to be able to, update. My asset, so. For example I'm. Going to at. Nate's. Google. Email if. You want to hit him up. And, as you can see I'm demoing from, another, organization, that's not Google so, once I click on add I'm gonna get a flag that. Just warns, me that this user is not in my same organization, just to make, sure. That. You, actually want to share it with them I, click. On save and confirm. The changes. And. Now my asset, is visible, or editable. To whoever, I gave access to and. This. Puts my asset in AI hubs ecosystem, it kicks, off the collaboration, process and, allows, other users to see share. And, update. My assets in just, a few clicks and not. Just that but, once, on this details page whoever has access to it can, download the source file that I uploaded or. They. Can choose to deploy, the notebook directly, in GCP, by clicking on this button which I will now show you. This. Is, redirecting. Me to, cloud, AI platform, notebooks, which, is a product that was announced in beta, back, in April and. It. Allows you to run hosted, notebook, instances, in the, cloud in a very safe way without, having to worry, about the underlying infrastructure. And, each. Instance, comes pre-loaded with, the most popular, libraries. And data science, and machine, learning you can also add your own libraries, you, can change, the environment, switch, up the compute power as you see fit and it's. Just a very easy way, to run. Your notebooks so, I select, that instance. Click. Continue, and this, opens. My notebook, on Jupiter. Lab and. In. The interest, of time, I have. Pre-round. The notebook before and.
As. I told you before this notebook. Builds, a binary classification, model, using XG boost it. Trains. The model which predicts, whether, how. Is under, or over. 160. K dollars. Which. Is around. 124. K pounds. If my, math does not fail me, then. You it, deploys, a model to the cloudier platform, and it. Uses the what-if tool which is a new feature them excited, to show you in a bit so. I really just have time to blaze, through snow. Book but, let's get to it I. Do. Some initial setup I import. Libraries, which as I said it's very easy because the instance is already preloaded with them, and. The. Second section I get, my data I. Prepare. It I drop some columns that are not that useful and end up with this information about, houses, that. Includes, square. Foot the, neighborhood. What, year it was built if it was remodeled. How, many cars it fits in its garage and, my. Label which is the sale price I. Do. Some feature, engineering I dummy encode the, categorical, columns and, I. Turned. This into a classification, problem by. Converting, the, label so the sale price into. A binary format so, a 1, 0 on whether, it's, over, 160, K I. Split. My data to train and self test and. Test. And train, my, model, using, XG boost resulting. In a model that's. 92%, accurate. Which. Will do for now and. In. This next section I deploy, it to AI. Platform, by, selecting, a project a bucket, and in. A, few minutes I've created, its, first version, it's, been, deployed. And once, it's deployed I can, use. The. What-if tool to interpret, the results and in, case you haven't heard of it the, what-if tool is a, visual, interface, that, helps, you understand, your data set and your, model results it's, very, easy to use as you can see we're instantiating, it, right, here with only a few lines of code we. Are inputting our. Test. Data. Selecting. A notebook and. Selecting. The target feature and, passing. In some vocabulary, and when. I run that this renders, the, tool that you see right here. We. Are first dropped into the, data point editor, which. Shows. Us our. Model, predictions, for each of the data points that we passed in so each of our, test. Data points and, in. This case we, see the blue points which are houses that our model predicts are over, 160, K while. The, red points. Are, predicted. To be under that price and the. Closer the, points are to. The middle the, less confidence, our model, is about, their classification. And a. Cool thing to do in this editor is I, can inspect. Individual. Data. Points so for example I just clicked on that and on. My left I can, see. The. Feature values and. I. Can actually, edit those. Values, to see how, they affect the. Model results for, example I'm going to switch, this up and going to say. That this house was remodeled, in a more recent year, I, then. Run, inference, on it and I can. See how this affects my model as, you can see my, model now is highly. Confident that this. House is over. 160. K and. This. Example is rather intuitive, but. As you, probably know, understanding. Models can get way harder, in. Real life examples. And it's. Very, convenient, to have a tool like this at hand, on, the. Second tab we. Can see how our model. Performs, and some common, metrics, we, can also optimize. For, fairness and, on. The last tab we can see how balanced, our data, set is and. As you can see there's a lot, to do with this tool a lot of options if you. Haven't, yet, I would recommend that you take a look especially. Given. How increasingly. Important, it is to really understand, our models, and make. Sure that, we are trying to decrease bias and just. Developing, AI responsibly. So. That's the, what-if tool. Now. Let's say I make some changes to this notebook and I, want to update this changes, in AI, hub, well I I'm. Going to show you a feature that, we're working on currently and. It. Will let me do this very easily. So. This, is the, notebook that I just deployed and. So. You can see we will soon have this. Plug-in in, Jupiter. Lab the. Bottom in the tool bar which. Will allow me to upload, the notebook to AI hub so it's just one click. One. Click and, this, redirects. Me to AI hub where. I'll be able to choose whether I want to upload. This notebook as a new, asset or, I want to update, an existing notebook. With a new version so. I select the notebook that I want to update. Click. Save and. A. New version of this notebook is created, I can. See it on the details page and I can share it as a. Previously.
Shown You, know. Yeah. So that, wraps, up some. Demos. So far we've shown you how you can, discover, a notebook and I hope you can also upload, your own share. It get. It running on GCP immediately, and once. It's there you can play around with it to discover, new techniques, plug. In the what-if tool and all, these tools really, increase. Collaboration and, smooth out the onboarding, process by, allowing, you to make. Your assets easily. Discoverable. And usable. Along. With the what-if tool which just increases, debug. Ability, transparency. And fairness and, to. Talk to you more about how, they use a I hope to, increase, collaboration, within. Their organization, I leave you with Richard hacking from, add sauce thank. You. Good. Afternoon so. If we could fleet weights the slides cool. Excellent. My name's Richard hacky I head up atlases UK, an island AI lab and it's a real privilege to speak to you today. So. I don't know how many people know about Atos we've, got over, 7700. Experts who work in data and artificial intelligence on, working. On projects for clients globally on, everything, from value, consulting, through to data visualization. And then, also over the last year we've created, five Google, AI labs where, we where we work with customers to help them Co innovate I look at what's, the art of the possible how, can data help. You to, get more value into your business. Within. The a I love so we have this proven, process that we use, we. Really start at the ideation phase so, there's lots of things you can do with your data but. It's really key to make sure that we can try and work on a you, know what would good use case be where it delivers value into your organization. And we think that there's the underlying data to support a useful exercise in, that so. We'll work with clients to identify a, good, problem area to look at and, then we'll then hold an AI lab session, which could be anything from one to two days and we'll, invite people from the organisation that are involved in the end-to-end process, we get multiple perspectives, of how that process work and it's like if you've got data that you can share with us and we'd bring that into the session and. We go through this three-phase, process so the, first process is really about discovering, what, can a I do so. Given the use case would look at how, do other organizations. In your industry tackle. That issue how, do other industries, maybe tackle, that issue to give different perspectives. And to give a bit of inspiration about what the art of the possible is and what AI technology, can do today, the. Second phase is all about design, so again we want to build upon how processes work at the moment so, we'd use a design thinking, approach to, understand, how, does the process work today what. Are the pain points so where is the cognitive load. Where people have to think a lot or we find lots of error rates in processes, and we'll all brainstorm on where, would good, good places be where AI could intervene and help, and, sometimes, that design phase we look at multiple solutions, which which might not be AI it might be a automation. Or an analytics type solution so again we can say is this a good problem for AI or is there a better cheaper way of doing, it and then. We do the define site so saying from, that design phase we've got a hypothesis. Given. Data we might have used in the lab we done some initial feasibility. Assessment, on do we think this is a good thing to do and how would it work and then on the define stage we then define a roadmap of like, how can we prove that what, the KPI so know whether this solution will work and as. Well as look at the technical, solution. Will, also look at what. The other things to do to get acceptance, into the organization. We find a lot of clients they can do the technical proof of concept, or the Minimum, Viable Product, but to get users to use it you need to look at that change management piece the business case the sponsorship, and we can plan all of that in and for, this to be successful for us it's, really key for us to be able to work at pace to, be able to build on proven components, used before so that we can actually spend the most you, know most valuable use of that time that we have together to, sort of see what the best possible way forward would be.
And. Some of the challenges we have when we're doing these projects with clients, are. Listed on the board here so the first one is about collaboration. So, although we have our UK, and I lab in London, my team is spread across the UK, and we, also work with God global colleagues in other countries, and it's key for us to be able to bring the right expertise, to a problem so. Again it's knowing who who's, worked on similar projects before and then being able to collaborate on one system in one way, even. If they are not physically in the same location, as us and. Best, practice is really a key point for us as well so. Again we want to be able to use the same tools the same approaches, where possible, to make sure that we're actually building. Upon worked on before and it's, really key for us to be able to quickly identify how. Is this done been done before in, our organization. What the lessons learnt that, we've had from previous implementations. That we're not starting, from scratch can build upon the, work we've done elsewhere and. The third thing that we find about agile projects, is that there's. A lot of experimentation, that we need to do and again, for a you know the amount of experimentation, that we're able to do within a time that we have directly. Contributes. To the quality of the end solution that we have so, both in our and our labs experience. Where we're doing the design we might experiment with different ways of solving a problem to try and work out what, would be the best way forward when, we're actually going to the delivery projects, and we're starting to actually use real data and, starting to put, it through more mature ways, of working again, we want to experiment and what's the best way of doing it if we augment in our data with other data sources does that improve the model doesn't improve our accuracy and, we need to be able to work at pace be, able to drive that quality. So. Here's an example of, some of the AI of. Our use of AI have at the moment it's on the notebooks that we use and. Again we find it really helpful to have all of our notebooks in one place one, screen that we use globally, so that anyone in our company, can see where, have you done things like this before, we're.
Able To tag on notebooks, on like the types of problem domain all the types of technical functionality, we're using so, it's very easy to search through and see, what, are the projects have we done that are very similar so again we can build upon previous, proven. Products and then learn about the limitations, of them and say what's the best way forward and one. Example would be I speaking, to a client on Monday and we're looking at doing some, work about predictive, maintenance on, smart fridges so they've got a fridge, they've got data on that on how its operating, and it's saying what can we help them predict when that fails so, therefore we can intervene, in advance we don't lose the expensive. Contents, inside. And again we were able to look at this and looking at the third. One down we've got PCL lab analysis, so that that PCA, means principal. Component analysis, so if we're trying to predict failure, which, of the components, and the datasets, that give, us the best insights, into when. A fridge might fail so it might be when we see variations. In the temperature that's, a lead indicator, saying that fridge is going to fail in two days time we've, got a proven model here as a starting, point so that quickly we, can start saying what, are the key features we need to look at so we can hone in quickly on one. Of the right elements to try and get that accurate, prediction, of when a fridge might fail. Second. One down from the bottom asset, location visualization. We. Did a similar project with another client last year and again we've uploaded that notebook into a I hope now so, again we've got something out of the box that helps us visualize. Some. Of that data in there so we're not starting from scratch go and build on that and everyone, globally in our company is contributing. To the same knowledge repository, for us, and. Again this is available to everyone with who's got the Google license within Atos so we have to build to this contribute, and build forward. So. Some of the benefits that we're getting the first one here is about quick onboarding, so of the, 5ai labs we've got two of them have only started in last three months in Dallas immune ik what, we found with using a I hope is that from day one these, new teams are newly formed have access to all of the algorithms. The datasets the models that we've that all the other labs have built previously, so from day one they can build upon stuff we've got and then contributes, that AI hubs going forward so not so he's not a new team working in a silo developing. New ways of working, we've. Got that central knowledge base so again it's very easy to search through all the different artifacts, we've got to see where have we done this before what's. Worked what's not worked and the. Third thing is about best practice, so it's helping girls to drive good, governance good, standard tool sets and approaches, so that we're, actually working, the best way so things. Like bias and fairness, is something that we hold, dearly so especially, when we're looking at sensitive. Areas with data which could impact our citizens, treated, we want to make sure that we're testing for bias in the data set but taste for biases, within the algorithm, and again we've got proven ways here of how we doing this we're not starting from scratch each time we can build upon standard. Testing approaches, to improve the quality and then the pace at which we can work. So. Looking forward we've been working with Natan and the Google team now for it for quite a while and again, these are some of the features I'm really excited that hopefully will be coming into AI hope in the near future, so, notebook reproducibility. Again. When you're actually developing a notebook. The underlying, infrastructure, might change by, being able to store some of the context, of the environment, with the notebook it, helps you to make sure that the notebook will work or I'll be able to debug it quite quickly on, the. Rich machine learning metadata, this, is all about as well as storing the data storing. Metadata about that so if we were able to store days for about what, dataset was used for training a model how long was it trained for what the limitations, of that model that means we're saying we're looking through a new project. We got three examples, of where this has been done before that. Metadata can, help us work out which is the best one to use for this model so again it's simplifying, the exercise. Of the data scientist to be able to get the right right, solution, first time and. Then. Data discovery, you know being able to extend the ability to search within AI hope to be able to look across multiple, datasets multiple models over public data as well again.
I'll Only accelerate, the, time so we've got one window to look at all these things. So. That's, my little bit if, anyone would like any more questions about this afterwards happy to speak after this session or I'll be on the a tile stand in the center of the floor and, happy. To take questions on using an AI hub effectively, or why draw now how, AI could actually help your organization, so with that thank, you very much I'll hand back to Nate. Thank. You thank, you Richard Mariana is awesome. Let. Me take this home so I want to talk a little about what's on the horizon for us right so and, before I do I want to talk a little bit about hay. Hubs, like current position and we're we're like focused on our main themes, going forward so the first is you've already seen. Primarily. Examples, of a notebook of how we can use, a hub, to store this notebook with, context, share it with others quickly, run it on GCP especially great if you guys are already using, AI. Platform, notebook service, but. You may be thinking like okay what. Do I use today you something like github or something like that for this process, and github, is great or other other get based source control systems we. Love we love using them and we use them as well ourselves but, there's a lot to be desired when it comes to data science, in. This area, one is that models, and notebooks can get very large and. And like source control repositories, will blow up especially as you start versioning, them I think we did some studies recently, and like, roughly. It takes about four months or five months when. We see a data science repository, to get to the same size or exceed the size of like 10 years of software, engineering repositories. With thousands and thousands of commits, and. That's because you're basically storing a whole new copy of that notebook or that model every single time you're versioning so you need a better system to manage these, that. Actually one. Is better, for larger, images, or larger objects but, also understands. What they are and the context, is so important, for machine learning artifacts. Is to. Get stuff is just a file and. To a hub it's a notebook and it's a model and which means we can show you detailed. Metadata and lineage about these items especially, if they're automatically, created, on google cloud zi platform, which. Leads me to our main themes for what we're working on going forward the, first is easy asset ingestion so, we showed you today how you can use our GUI to. Upload assets, but of course for things other than notebooks, in particular, that's not going to be the best method, for actually, adding my assets to a hub's ecosystem, and, so what we're really focused, on is providing a programmatic interface for, you to build a hub into your existing workflows. This. Could mean one. Actually. Integrating. With your git based workflows, but also, providing. A napi, to use, as well so I would stay tuned on that item this, could be this, is really really great for particularly. Storing, models and pipelines. Second, is greater, integrations, with the. Rest of the AI platform, like, I said earlier one of our key. Principles. Here is that all of your assets can really quickly be deployed and run on other, cloud, platform, services and, vice. Versa so we want to make it really easy for the things you create, on the ed platform to, automatically. Be available in a hubs ecosystem, for sharing with, all of their detailed, metadata, and. We also want to make it very easy for any asset in a hub to automatically, be deployed a model, for batch. Prediction. Or just, for testing on a new data set so, everything should be aspirationally. One-click, deployable in, a reproducible manner this. Includes some of the notebook reproducibility, work that that Richard alluded, to which, is basically the idea that if you're running your notebook and creating your notebook and offering your notebook I should say on, Google. Cloud platform and, storing it within a I hub we should be able to maintain the context, of the environment that was running that notebook and make it much more likely that you can open that notebook, up in a compatible environment. The next time you try so, you don't run into the situation that I hear, about all the time where again, six months down the line you, can't figure out why this thing that you created isn't working as expected and last. Is enhanced collaboration so we, showed you today how you can you can share things and you can edit and update assets, and update their metadata but. Collaboration just starts with sharing right once you put things in people's hands there's, a whole host of collaboration, workflows, related. To the machine learning model. Life cycle development, like as things, go through stage gates and approvals and evaluations, that, we will be building into into, the product as well so. Super, excited about all this, to.
Make Machine, learning a much more collaborative effort. And trying. To reduce the redundancy and, increase. The visibility traceability. And, auditability. Even, of all the machine the new artifacts that you guys are creating in your organization's, and. That. Wraps it up. Hey hubs in beta today. Please feel free to jump in and check it out, any, organization. That uses G Suite can automatically, from day one share assets, across, their organization completely. Free, if. You're interested, in the notebook that Mariana, walked through in particular, the what-if tool which is an awesome tool for explain, ability that. Notebook is actually available in, hubs. Public catalog so all you have to do is search for what-if, tool extra, boost or something like that and you can you can grab it run that right. Away on. Our home page we have a link to a community which you can join where you'll see some regular updates or chat, with us, raise. Bugs or feedback and, of, course any of the upcoming features, that, you may have heard about today or, just ideas please. Please feel free to reach out I will, hang out right here after the session and I'm sure I'll get kicked out and I'll we be waiting outside for, for. The rest of the afternoon if you have if you have questions or input, thank. You again for coming it's, been an honor to be here and, I'm looking forward to hearing from a lot of you in the future thanks, again. You.