The DataHour: Google Cloud AI/ML

hello everyone good evening and uh welcome to the another session of data so we are thrilled to have you here in uh in this evening uh for an Interactive Learning session my name is along with my friend Ram goswami uh we are the part of a data science team of analytical Vidya and I will be the moderator of this session so for those who have joined us for the first time a brief introduction about the data session is with the intent to make learning data science and more engaging to the community as you know the analytical Vidya is is a biggest data science community so we begin with a new initiative data which is one hour dedicated to the data data is our is a series of webinar led by a top industry expert where they teach us and democratize data science knowledge so moving forward uh so now today our session is about Google Cloud Ai and ml so in this data session uh Mona will cover Google Cloud AI tool which are armed with the best of Google research and Technology to focus exclusively on solving problems which matters so in this data we will cover Google Cloud vertex AI platform which is an integrated machine learning platform for developer and data scientists as well so we will also see a NLP API document and Ai and Enterprise transaction have in action right so uh before we uh kick things off and I hand it over the session presented Mona so a quick recap of the few things uh we are recording this session and we will make the recording available in few days on our YouTube channel please use the Q a section for asking any question you might have during the session as the data are progressed towards the end we will do our best to answer them right so and the last uh also we will share a feedback goal towards the end of the session and which I request everyone to participate in so now uh on to our speaker so in this session uh our data we have Mona with us she's an AI machine learning customer engineer at Google right now and uh Mona is then also an author of NLP book natural language processing using AWS AI service she is highly skilled IIT professional bringing more than 10 years in software design development and integration across diverse work environment before joining Google she worked in Amazon web service as a senior machine learning solution architect her role is to ensure customer success in building application and service on the AWS platform she has also published 17 blocks and on AI and NLP on AWS AI Channel and research paper on AI power search solution which is also published by Amazon science and also got runner-up in a triple AI conference she has a 14 certification and currently working on writing her book on Google Cloud professional ml engineer certificate study guide okay Mona you have achieved a lot seriously it's been honor to have you in this uh in this session and uh so guys uh I will uh I will share the LinkedIn profile of Muna kindly do connect and uh with Mona and for the future guidance so over to you Mona uh the virtual stays all yours uh yeah thank you very much RAM uh let me share my screen uh for presenting my content let me know if you can see the screen foreign I hope everyone can hear me okay and see me fine and then I'll proceed okay I assume it's yes so this was our introduction thank you ram um so this is just about me that I'm a customer engineer here at Google and also author of this book and why data hour and why I was interested to do this because it's very very similar to uh what is my uh previous book about it was about to empower every data scientist and ml engineer or anyone who wants to make a career in AI with no machine learning experience uh by writing this book so we we presented lot of low code low code Solutions and especially with little bit of python understanding uh you can get started with your ml career journey journey however today's session is about Google and why we'll cover these four four topics and we'll also see a lot of uh demos and live Demos in in this entire presentation and if you have questions I don't I cannot see the question so um if Ram can tell me if there are questions he can stop me and I can take a pause and he can tell me the questions I'll be able to answer it or we can take questions in the end whatever works so the four things we are going to cover today is why Google for AI the second thing is uh Vortex AI platform and we'll talk about what is interesting things in vertex Air Platform and how how it is different and how it is going to empower you uh we'll see uh walkthrough of uh vertex Ai apis and we'll talk about what is automl and how you can use automl and lastly uh we will also see some demos of AI Services which means that um some of the services built by Vortex AI which is uh document AI service uh which helps you automate all your documents and also Enterprise translation hub with that let's get started so some of the things why leverage Google Cloud for AI uh before that uh I hope you all know what cloud is and what is Google Cloud um can we do like a quick show of hands or poll to know that you know this you how much familiar you are with Google Cloud okay I assume that you all have some basic understanding of what cloud is and what Google cloud is and Google cloud has tons of services today we are just going to focus on AI specific services so some of the ways uh why how you can and why you should leverage Google cloud is to scale and when I say scale means that you get instant access to thousands of machines such as gpus or tpus with Google cloud and tpus are very very unique to Google Cloud uh what does TPU means TPU means tensor processing units and our TPU with TPU you can run large tensorflow workloads on Google cloud and they will give you the Speed and Performance and as well as scale you need which you cannot achieve in your normal system the second is speed we already covered speed with scale they both both the words come come together uh so we covered that cloud tpus and what it is and how how Google Cloud can give you speed uh third point is quality and how you get quantity is uh we have pre-trained ai Services which we are going to cover soon which solves basic business needs with highest quality so you do not have to reinvent the wheel so uh whenever uh when I started with machine learning uh I would go and you know find a lot of data first and then I will write a code to solve a problem right uh but with these AI Services it's very very simple you just you already have everything trained using Google's uh machine learning and data and what Google has Google cloud has done they have exposed an API to you uh where they have solved problems common to machine learning and you just send a request and you get a response back it's very very easy to set up and it's very easy to use and it can solve many common business AI problems which we'll cover in a little bit lastly customization so ability to customize um these pre-trained AI building blocks using transfer learning with your own data set is the capability I'm talking about so you can use cloud automl and ml engine to customize models and we also have advanced Solutions which is uh based on these customizations which is document Ai and Enterprise translation hub so this is uh to say that there is AI opportunities in every industry be it retail be it Healthcare um be it Financial Services be it media entertainment or public sector so or a startup so you there is AI opportunities and AI use cases so these are use cases so from demand forecasting to Telehealth to money laundering so it can be n number of use cases so it can be there is AI opportunities in every Industries and uh so that is why we created a platform which can cater to all these industries as well as all these users and the users I mean that in any kind of ml life cycle it's not only one person if you're a data scientist you're not only one person doing the job it's end-to-end machine learning and how it begins like you have you have to gather requirements first and uh most of the time is spent on identifying whether it's an AI problem or non-ai problem because today everyone wants to do AI they get very excited and uh first thing you should always ask why you want to solve the problem if you can solve it using simple non-air techniques I would say go for that because AI is doing artificial intelligence is really really hard and why it's hard because it requires lots of data because your model is going to be as good as your data is and also it's an iterative process it's not just uh somebody hands you the data and you're done you will train your model you will tune it with various hyper parameters and that requires a lot of compute setting up infrastructure also you would like to after you've done your model you would like to deploy that model in production and even that involves a lot of infrastructure heavy lifting because you will get a server first you have to go and ask for a server if you are productionizing it right and getting a server usually takes five to six months if you are starting your own business or or once you set up your server you have to think about scaling it right so today you have scaled it for thousand users and tomorrow there is some sale there is some Diwali a sale going on uh you will have 100 000 users hitting your endpoint and that is totally not going to scale your model and your model uh in production will fail so you have to think about scaling in those scenarios and also you have to think about how your development team and there's a lot of collaboration between all these users right when when we say that machine learning there's a lot of collaboration between a data scientist versus a ml developer versus an ml engineer to productionize your end-to-end code so it's not just one person doing everything and that is why uh we have created uh Vortex AI portfolio uh and which is called Cloud Ai and industry solution portfolio which consists of uh vertex AI pre-trained model and vertex AI platform so vertex AI platform we'll cover in details what all it consist of uh so it has two components and it also have the AI solution suits so when I say AI Solutions it means that you can think like software as a service it's already built for you uh you do not have to build it from scratch you can come and directly use it and some of these Solutions mentioned here is uh contact center AI which consists of virtual agents chat Bots uh so it's completely a solution to revamp your contact center so any industry can come and use it if you are Consulting any form you can recommend this to them document AI is one of my favorite I really really love this product uh and what this document Ai and we'll see this today in demo so document AI solution what it does is uh you upload any document it can be uh you know handwritten notes it can be any language document it can be uh written in any language suppose in Hindi uh and it can be uh any type it can be your pay slip it can be your invoices suppose you are student and you're submitting some research paper and you want to see what is inside you can uh use that research paper it can be in PDF or it can be images uh so you upload that to document Ai and you get a response back and in that response there is lot of lot of data which can be structured so it's a product which can use to convert all unstructured data into a structured document the structure structured information and we know that how important is structured information for us so you on that structure information you can build insights on top and you can quickly analyze it like who is doing what suppose if it is a procurement document or an invoice document uh you if you are able to extract these meaningful uh insights from it you will be able to know when was the procurement happened when was the invoice sent and who lacked like who which all persons are lagging in in the payments so all of this can lead to a lot of discovery and Healthcare Insight is another uh solution which use uh which is mostly focused on Healthcare products I'm not going to cover that uh today uh mostly we are focusing on pre-trained models vertex AI platform so in pre-trained models we have variety of pre-trained models which is uh trained using Google's data set and we are making it available to you and the problems solved is speech Vision translation language and structured data and with vertex AI platform we provide the notebooks Jupiter notebooks for you to get started and we also provide vertex AI training prediction and ml Ops will cover all of this in future future slides sorry to interrupt a few of the members are saying sorry uh few of the members are saying volume is level is low uh can you level up your volume right a little bit yeah sure is it better now okay yeah carry on yeah it's better okay so sorry about that uh so I I was just I just covered uh for those who cannot listen to me I really apologize I just covered uh why you would choose Google for AI why AI is really really hard to do and we introduce vertex AI platform which covers pre-trained models which we are going to talk now and we'll also cover vertex AI platform notebooks training and all those cool features and I hope you can hear me okay if not you can let the moderator know right now so talking about vertex AI pre-trained models uh which is all available uh so we have created pretend models for solving common machine learning problems for example we have Vision AI Envision we have Vision AI uh which can do a lot of things such as object detection uh out of the box so that you do not have to come up with thousands of data set build a training algorithm and then think about hosting it all of that we have already solved that problem for you and we are providing you an API and we'll see that how that API works today in language we have a lot of interesting uh Solutions we have translation um I'm not sure who all have used Google translate but whenever I go on vacation I rely on Google Translate us English to Spanish Spanish to English like you have tried like you can try Google translate anywhere in the world you go and Google this translation API uses the same API which is used in uh Google Translate so you can also use automl translation to uh you know do some transfer learning and make your predictions specific to your needs so you can take an example uh suppose uh you want some words to be translated unique to your language or unique to your business you can do that using automl we also have natural language apis which can do a lot of things such as it can do classification out of the box it can do entities it can do um it can also classify it out of the box the Google does that and you also have ability to do automl natural language where you will do a named entity recognition with your data set with your business needs and you can also do custom classification based on your specific use cases and you can classify documents you can classify invoices uh all with this automl natural language talking about a conversation or we have dialog flow which is uh the chat bot which enables uh the contact center AI service we uh so we also have speech to text apis which will convert speech into text and vice versa and for structured data suppose you have lot of structured data and you still want to you do not know what to do with it you can use automl tables we have a lot of new things coming uh time series insights API vertex AI forecast for for forecasting with this understanding let's deep dive into uh what xci first so why customer choose vertex AI uh there are four reasons to do that the first is uh unified data and AI platform for all users to accelerate time of value second is ml Ops because ml Ops is really really hard like managing if you are an ml engineer you would understand that it's not easy to manage and Governor distributed large uh machine learning environment uh so Vortex AI enables end-to-end machine learning Ops third is open and scalable AI infrastructure and when I say open and scalable AI infrastructure suppose uh you have you are running a Pi torch code or a tensorflow code you can still use vertex AI because it's open source and Google supports a lot of Open Source framework and lastly state of the RTI we will deep dive into each one of these here so talking about unified data and AI platform what does this mean this means that Google Cloud users benefit from an integrated data through value journey and what does this mean is that our our vertex AI platform is deeply integrated with our data analysis platform which is bigquery for those who do not know what is bigquery uh or a bigquery is serverless data warehouse provided by Google cloud and it's an amazing product so vertex AI workbench the manage workbench will directly integrates with bigquery so if you have data in bigquery you can directly use it in vertex AI or you have data in uh storage Google Cloud Storage which is similar to Amazon S3 you can directly point to that data in vertex AI workbench and start using for your training purposes so this and why this is important data value Journey because we realized that in for machine learning data is very very important and 90 of time is just spent in having clean data set data labeling is very important you need to have a label data set for your model training and that is why this and and the third thing is data plus value journey is that you can bring data from any any platform so it doesn't have to be on Google Cloud suppose your data is in Amazon S3 or in some other cloud provider or you have data stored in iot devices social media or it is on-prem in your own local uh computer you can still use all these supported data sources for your ml development using vertex AI so ability to get data is also very important you are getting the data and then you're using uh vertex a platform to training on it talking about second point about end-to-end machine learning Opera operationalization which is called ml Ops ml Ops is is um A New Concept and what does this mean is that uh suppose you are training a model right so it's not always that you will be training and then you are done with it it is a process where you might need to retrain your model suppose your data changes uh in production right um let's take interest rate it is very volatile you have trained your model A regression model with a uh with suppose or set interest rate 3.5 or so take it six percent and with inflation now your interest rate has become eight percent in in production so your model won't perform and that is why you would need to retrain your model that is just one example there can be retraining in multiple cases you can be your model has changed or your data has changed uh and to do that it's not easy right uh retraining is not easy you have to again set up infrastructure you have to figure out what parameters changes and that is about training and experimentation then talking about model deployment you also have to continuously monitor your model in production for those changes about data skew or concept skew so that is why uh vertex AI has created an end-to-end environment to yeah that is why vertex AI has created an end-to-end environment so that you can you do not need to worry about uh setting up end-to-end ML ml Ops and lastly the piece is about a Model Management and governance and why this piece is very important when you work with multiple identities like we say we covered initially that vertex AI supports multiple users like if you are a ml engineer versus a data scientist and and why it is uh and how it does that is using vertex Air Platform using uh Model Management and governance so you have a model registry where you'll go and register a model uh so that you can track the lineage right who trained with what data so model governance is very important when you are setting up a secure environment and compliance is also important we also have uh feature stored model feature store and what feature store does that you create a feature suppose I am a data scientist and I'm working with your team you are also a data scientist we all will be we are working on same problem we all will be doing the our own feature engineering and we all will be converting the same features using same training code and it is very it's not scalable and it's not this is very repetitive repeatable process so that is why vertex AI has created a feature store where this is one time process you will run a feature you will convert all these features and store it in feature store so that complete team can use it while training as well as during prediction so some of these Advantage with MN Ops you get using vertex AI moving on to next last uh third point is open and scalable AI infrastructure so vertex AI I already covered this data value it any data source it can be bigquery cloud storage it can be file store it can be multi-cloud data right any data source any framework so all of these Frameworks are supported by vertex AI platform and it can be any hardware it can be CPUs gpus gpus so that is how you it can accelerate the velocity of models to production by providing these all flexible environment lastly talking about state of the RTI what does this mean what does this mean is that from open source to so this is all blue is in open source from Hadoop to tensorflow to tensorflow Extended Cube flow all of this is open source work done by Google so from open source to Google research to gcp solution if you see the timeline it takes less than two years so what I'm trying to say that Google cloud is dedicated to Bringing open source source Google research into production as soon as possible and that is how you are having access to state of the art AI starting from bigquery to Cloud tpus to automl to uh Nas and to vertex a matching engine all of this is has come from uh Google's open source research and also tabnet we are converting all of this into Google Cloud products as quickly as possible talking a little bit what is neural architecture search and how we say state of the art Ai and how we are implementing it in 2016 um neural the paper on neural architecture search was published and the idea of using a recurrent neural network to compose neural network architectures so once this article was published the in 2018 it was adopted adopted by waymo and other alphabet teams and by 2019-2021 uh we launched vortex ainess and one of the customer Qualcomm is uses this product and some of the advantage of using Nas is that they were able to optimize their AI models in weeks rather than months so it was a game changer for that business some of the customers who are using um Google Cloud AI is Iron Mountain using it for automated document processing eBay is using to transform customer experiences uh Rolls-Royce uses for autonomous vessels and HSBC is using uh some of the AI techniques for risk management now we will cover uh some demos and walk through of our AI apis so there are two type of AI building blocks when we call about AI services uh the first is pre-trained models and the second is custom models so custom model is automl right uh so let's see some of the differences between these two first is uh when we say pre-trained models you do not need any model training it's already uh trained and you can leverage Google's tried and tested data set and it's very very ideal for common use cases talking about custom models custom models have provide you high accuracy with models built using automl technology and it leverages domain specific data set plus Google's data set and it's ideal for specialized use cases or custom use cases foreign so this is a quick walkthrough of how automl will work uh you will bring your data set and the advantage is that you do not need lot of data to get started like I think for uh some of these examples you can get like 100 samples 200 samples and you can still get started rather than you know having large training data set to get started so you bring your own custom data set um you do automl training and you can you have choices to deploy um and you also have choice to serve it as a rest API you can download this and deploy it on edge devices too which is called automl Edge we covered this already let's talk about NLP API and what it does and we'll see that in a demo so it does these four things out of the box or it will extract entities it will detect sentiment uh analyze syntax and classify content I will take you to a demo you all can come and use this uh try it demo uh these links are very much accessible and that's what I'm going to show you today so just try the API demo and how it works is you can grab a text let's grab a text about analytics since we are in data hour let's find out how it's going to extract meaningful information about this analytics with you let me go to the website and grab some text okay okay let's just copy this uh oh it's not copy let's try to get it from Google then so you come to this page and you do try the API and you can paste any text here let's analyze what it says yep and so once you click on analyze you start getting interesting things about it this text so this is a small texture but if you if you have a large number of text or suppose a review right and you want to know the sentiment of the person writing the review whether it's positive or negative you can quickly come and see the sentiment so for this it should be uh neutral right because it's a just information about a document so let's see what you get out of the box so you get uh it is identifying analytics Vidya as a person and this is where automl can come into play right you want analytics Vidya to be identified as organization so sometimes most of the times out of the box detection might not work well so this is where you will need automl that is the classic case of automobile training for identifying analytics media as an organization rather than a person uh but it did classify other information correctly uh like data data visualization it has to do with machine learning artificial intelligence business analytics that is all true so for sentiment it gives you two types of sentiments uh overall sentiment of the document and also entity level sentiment so you get entity level sentiment which can be used for use cases where you want to create Predictive Analytics in based on the sentiments it does syntax analysis and this syntax analysis can be used as part of speech tagging for creating about model right uh so this this can be out of the box used and lastly categories uh this is very very interesting uh it was able to identify analytics Vidya in jobs and education which is right computer science and business and Industrial just by a few few texts which I haven't even seen and for complete list of categories you can come to this page and check it out so uh it has all of these categories which does out of the box which the API uses out of the box let's move on to so we covered NLP API and these four things out of the box entities sentiment syntax and categories so some of the use cases you can use is uh sentiment analysis with your dashboards right so you can use this sentiment API for use cases where you are getting data from you're analyzing data from a dashboard suppose Reddit sentiment analysis or your check or your you have a business right and you want to see what uh people are responding to your business whether it's positive or negative right on in real time you can use these apis in an application where you can know the sentiment in real time and react before the damage is done so this API has a lot of use and it's a very powerful API at a very powerful way to uh innovate your business and think out of the box and entities have very wide use because most of the time you would go with custom entities right you would know a named entity recognition is is an area of natural language processing which is very very tough to do right so with these entities you're getting out of the box entities and using automl most of the customers I've seen they use automl because uh they want custom entities unique to their business suppose like we saw analytics Vidya was out of the box or identified as a person but it should be an organization so this is where you will do custom entity training you will say that hey call analytics Vidya as an organization uh in in the data set in the your training file and provide that to automl and these apis automl API will start identifying that as automl as an organization so this is what automl natural language is uh upload and label text so you will upload uh for your text it can be Sports lifestyle Tech any any type of labeled data and then you can use to train your model and then you can evaluate it so this is about translation API how translation works and how automl custom models works uh so out of the box so this is a Spanish translation and this is Spanish to English translation you see so custom model and the Google nmt model is our API which you get out of the box so Google Out of the Box gives you this translation while with custom model qualifiers and graders like there is some difference like you are making it unique to your use case that's what uh is automl making things unique to your use case now talking about Vision API Vision API does a lot lot of a lot of things it can classify content with predefined labels it can do OCR detection with 200 languages it can detect objects detect Brands and product logos um you can find similar images on the web which Google already uses this you can detect popular places and landmarks detect facial faces and emotions identify image properties and we also have video API which supports video uh object detection in video as well and track object tracking in video um so let's see this API in action to understand what all these features are and all of these apis comes with their drag and drop try it page and this is all free to use you can come and try it before getting started then you can drag an image click I'm not a robot and this is what you get out of the box so uh you can come here um upload any image to get started and you can see that it is doing Landmark detection I really don't know where this place is but this API is extremely smart and it is telling me it's Jakarta History Museum with 89 confidence and it is also uh giving me the exact Square where it is with the Google Maps link which is pretty pretty remarkable very very remarkable feature moving on to faces detection so it gives face level uh confidence that how emotion uh level confidence how How likely and unlikely the phase one is and How likely and unlikely uh like Joy sorrow anger surprise for phase two uh with the confidence score respective confidence score talking about object detection it was able to identify what all objects are there in this image starting from bicycle person uh hat all of this is has been identified from this image uh talking about label Rejection it does out of the box labels for these categories bicycle tire building mode of transport fun travel City cycling street so all of the labels associated with this image is here talking about OCR and text detection it was able to identify this texture which is in a different language I do not understand so you see this it's the text is also detected and uh it also determines the dominant color from the image and safe search is about whether this content is safe for search or not all of these features is out of the box and you can use custom Moto ml to customize some of these features based on your use case moving on to so we covered Vision API and we covered the demo [Music] talking about automl Vision suppose uh we saw what we saw was uh the API level uh use case suppose uh first I would recommend drive with the apis right if it doesn't work for you then go for automl there was a post that did not work for you and now we are doing Auto ml how it's going to work you will have to upload the images and then you can use vertex AI data labeling service to label the images so vertex AI comes with labeling Service as well so you do not have to worry about how to label the data right you can just upload suppose I I love handmarks I will upload 100 images of the Handbags and I will label it hand back hand back handbag all of this handbag and then I will start the training with one click which will automatically pick the best model you don't have to worry about writing any code and once the model is trained you can upload unseen data or a test image to evaluate the model so you will get the F1 score you will get everything you will get Precision recall and your model will be classifying handbags for you we also have Auto ml Vision Edge uh and Edge is uh what you get out of the box is you can export the model and you can you get a tensorflow model after uh training The Edge model and you you can also export the container and deploy it anywhere so you must be thinking that oh it is automl I cannot change the code you can still change the code you can still use it and how you can do that once you export the model you get uh you have options how you want to export it you can export as a export it as a container and deploy it anywhere you can export it as a tensorflow lite you can export it as a tensorflow.js for browser or you can explode it export it as a as a normal tensorflow library and once you have the code you can add additional layers and you can retrain the model if you are a data scientist and you do not like automl you'll be like oh I do not like automl because I do not know what's going on inside and another interesting thing is all these features comes with explainable AI integrated and what is explainable AI is a way to debug the Deep learning models and to see which features are attributed to this explanation so all of this comes with explainable AI integrated when you start training so you will also have an explanation saying that which feature contributed to this training so which is another cool uh cool uh I would say integration of vertex AI with explainable AI vertex explainable AI talking about uh Google Cloud Speech to Text uh Speech to Text comes into uh 120 plus languages and it can be real time or you can do it on-prem and uh I'm not giving a demo speech to text today but just to let you know that that is another service we all use even if you are using have you if any one of you have used Google meet that use a lot of speech to text in real time text to speech it has 40 plus languages 220 plus voices some of the unique features are it has human-like speech and it is one of a kind voice features which is unique to uh this product where it doesn't have to sound like oh I'm talking to a machine moving on to walk through of AI Solutions and I will cover document AI first and uh so document AI solution users we covered Vision Ai and we covered NLP API right you can think document AI is a combination of vision and NLP API to make the information to make the the data an unstructured documents useful uh how it works is the first passes through vision AI under the hood you don't have to worry about all of this I'm just giving you the details how under the hood document AI works so the first pass is through vision AI to extract the OCR data from highly unstructured documents it can be PDF it can be images it can be word and then use NLP API to understand the meaning of it and then you can either store it in bigquery or any kind of database data warehouse to make it useful so let's see a quick demo of document AI again you can come here um for the try it document AI page I'm going to upload a invoice so this supports so many things like you can come for General you can use invoice parser payslip parser so what does this mean is that these parsers or these uh options are specialized for those type of documents so it will give you better accuracy when you choose these parcels also you have ability to create custom parser suppose none of these matches and you have a specific form you need to be uh extracted right that need to be extracted you can come and create a custom parser or a specialized parser it also supports human in the loop and a lot of other features you see that out of the box it's able to detect key value pairs so invoice number date of issue build to you can quickly search on something if you want and then it also detects the table in the text right it has two tables it detected it beautifully and it also gives you all the OCR text and a Json output so this is the request URL request and response and you can download this so this API can be integrated anywhere in your mobile device or web application and it can be used to get your data converted into a structured highly structured format so we have like 10 minutes left um I will move on we have few topics to cover these are some of the resources you can use to get started with document AI I'm going to share it later and I know we did not cover a lot uh there is so many features of document AI you have custom parser you have a specialized parser you have human in the loop uh which we haven't covered today but just to let you know all of these features exist talking about a Google Translation Hub it is a Enterprise ready uh self personal translator and how it's it's like very very easy to use and it's uh it comes with multiple of language support so you can start with any kind of these documents it can be PDF document PPT document and you can quickly translate with one click uh self-sub training and you can also do customized translation with automl customer translation which we covered before so I will give a quick demo of Enterprise translation hub okay so I am in my um this is my Google Cloud console for those who have not seen it before and in Google Cloud everything is in project and this is my organization and this is these are my projects so I'm in my project and I will go to translation hub and in Translation Hub you have ability to create a portal where we will log in and upload the documents to translate I have already created a portal um I will just grab a link to the portal URL to the portal you also can add users to the portal like I have added myself as a user here uh so this is my portal and let's go and Translate so I have this Google nmt model which is out of the box and you these are some of my custom trading models you see English to Vietnamese custom trained models I'm using uh Google nmt model and uh upload a file I do not have anything so I'll just translate the invoice so I have this invoice and I'm going to translate it to let's pick any language I'm just picking languages out of the box I do not have any preference do you see that it started translating into these three languages uh and it's already completed so let's see the translation so you can also see the translation uh here you can see that it is already translated into this language which I really do not understand but this product is very very easy to use and it works beautifully okay moving on to um so we have 10 minutes left I will take a pause and take questions is there any questions here yeah there are a few questions in the Q a sections uh hope you can see those questions uh where can I see the question oh yeah uh will Auto ml dry data science jobs uh I would say automl does not dry data science jobs it enables data science jobs because I work with a lot of data scientists in um public sector space and they are using automl to uh they're not using automl to you know just drive their job they're using it as a as a tool for a faster Discovery and for faster innovation so if you want to innovate something really really quick because time is is of essence right uh and what you can do now you cannot do later so something which can take ages to do with auto ml you can do faster and better and you can I also covered that uh if you're not happy with uh what automl is giving you you get the code you can import the model itself right so you can add your data science experience suppose you get a tensorflow model back you can add extra layers on top of it so your data science job is not drying out it is well uh I would say supported I hope that helps Arvin okay um Roya has a question what is Vortex AI Vortex AI is a platform provided by Google Cloud which has products for end-to-end machine learning so it you can think like three layers it has Auto ml it has pre-trained apis and it also comes with workbench uh which we didn't see today it comes with workbench which similar to Jupiter AI notebooks which you're using and also comes with a lot of features such as uh model training so you can use vertex AI to train any model you suppose you have a model in local environment uh scikit-learn model or XG boost model uh and you are limited to train uh like you have you cannot train it on your computer because of lack of compute and resources uh you can come to vertex Ai and upload the code and use vertex AI training to scale your model for any kind of GPU CPUs you can do distributed training and you can also use vertex AI for prediction suppose you already have a trained model but you do not have resources to set up infrastructure you do not want to buy a server which will take ages and then uh you know set up infrastructure for training you can come to vertex Ai and you can do vertex AI prediction using vertex AI prediction uh feature which will automatically take your code set up a training for you and give you an end point where which you can embed in your website and uh and it is scalable also because we are taking what the Google Cloud takes care of the scalability I hope that helps it's a short answer it's it's it's a very it's it has a lot of features but I hope this helps what is the accuracy for document classification uh where is the question where did I go it might be the third question just yeah yeah Naveen has a question sorry you're saying something uh no no just carry on carry on okay so what is the accuracy for document classification with vertex AI uh you mean uh document classification with our NLP automl or uh document classification with AI services uh I would I would think that it is you're talking about the NLP document uh so you're talking about document classification using automl uh classification so the accuracy depends on on the data right and you also get F1 score and also you also get the recall value to improve the accuracy if you're using automl so there is not a set answer to this that's what I'm trying to say it totally depends on on the document but there are ways to improve the accuracy if that's what you're asking and I can share some links later how is our data secure when combined custom data with oh okay I do not see the questions anymore okay how is our data secure when combine custom data with good data uh so the question is about security uh I would say that you have choice to use encryption while training so vertex AI you can put the data in a VPC yeah you can put your training in a virtual private Cloud when uh spinning up Jupiter notebooks uh it it will be in your own secure environment for uh and Google Cloud uses encryption in transit encryption and rest so I can talk about all those ways to secure your data if that helps foreign does Google take custom data what does this custom data means I really do not understand but Google has ability to get data from uh any Source uh it can be uh from any other cloud provider it can be Amazon S3 it can be on-prem system you can bring data from anywhere to vertex Air Platform and train your model uh feature store generates feature or just one stop place to find all the features that are pretending to the project features stored do not generate future I would say it is a place where you can store the features after the training and uh after feature engineering and you can use that for your prediction and you can use it across teams so the idea is that you do not have to do feature engineering multiple times free trial has limitations like 30 days yeah so Google Cloud does comes with a free trial and you get 300 of credit to get started and some of these drag and drop try it apis I shared you can just go ahead and use it there is no limitations uh and you do not get to you don't you don't even need a card to register or use it you can just go to that link and play around with that so it's as simple as that how to make unstructured data structured that's a good question how you would do it is using uh so if the unstructured data is is in PDF or it is in images you can use a document AI solution to X convert the data into a structured format which we saw in a demo and you can store all of the structured data like key value pair you get all that key value pair right so you can store all of this into uh in any any data store you would like what is OCR text that's a good question Roya OCR text is OCR means Optical optical character recognition and it's an old technology previously it was not powered by Machine learning right uh it was a template based technique where uh you have a PDF you have a form and you would like to extract that information like suppose you scan something right you have a PDF or you have your notes you have uh you go and scan those things before you exam right uh when you scan it it is converted into a paper or a scanned format and then you want to digitize it right so in order to digitize it or convert it into electronic format and then extract the details from it that's what OCR will do for you uh there are more questions coming is there a possibility okay where I was do we have a demo for no we haven't covered automl for today but in future we can and I can share some links where you can get started with automl what is free to your Google ml so there is no Google ml feed free tier Google cloud has a free trade where you can go and create a project and use 300 up to services so that you can use uh Sashi Kumar has a question is there any possibility to compare the data with different AI models for classification and is also any validation for the model selection uh so we covered neural architecture search right what Nas does is compare various models and pick the best out of it if that's what answers your question but I'm really not sure if I understood your question correctly uh there are more questions compare data okay our linear regression and regression trees same I really don't know this uh because I did not are you talking in general data science because I have heard about XT boost which uses trees and linear regression is a standard way to rule linear regression I'm not sure what this question is uh automl translate in end-to-end data project compile the gratify ingesting uh I really do not understand what you're trying to say monali can you if you if you can type it again I would love to answer that briefly explain neural network model prediction um I I did not cover neural network model prediction I just want to specify today what we covered is Google Cloud vertex AI platform prediction and how that works I can specify and I've also covered this in my upcoming book and I've also covered neural network model prediction in my upcoming book on Google Cloud certification so that will help because it's not a short answer to explain how this Google cloud has advantages over nine platform okay so uh I have used nine two Naim is another product and it has limited capabilities right you cannot use nime for auto scaling training your model or to do ml Ops right while Google Cloud comes with a lot of Google's research and capabilities to support uh be it speech to text apis be it uh translation be document yeah it has wide wide applicability so that's why it's it's very very different and I'm Naim has very limited capabilities is there any payment for vertex AI you mean to say how billing works uh it is similar to Google Cloud billing pay as you go service so it's your only bill for the use you're doing okay there are more questions so we are ahead of time I will take a pause I know um I have to jump for another call is I will be looking at into these questions and answering over email or you can connect me on LinkedIn and there I'll be able to answer these questions because I'm really really sorry I have a hard stop today okay I guess uh you have enough questions answered so so first I I would like to share a feedback from uh hope everyone will uh see it and fill that form and thanks a lot Mona uh it was a really wonderful session and uh on the behalf of analytics Vidya sure I'm really sorry if everyone can uh you know if you can share the questions to me I can answer it uh like I can answer all of those and send it back to you if that helps or they can connect with me directly on LinkedIn and I'll be able to answer them definitely like we have shared your LinkedIn profile uh they will connect to you on LinkedIn and uh probably they will ask all the questions related to the topic there okay perfect thank you thank you so much thank you thank you Mona for uh for this wonderful session