Vertex AI Key Features

Cloud Storage/BigQuery

the first is cloud storage and BigQuery so cloud storage is Google Cloud Platform (GCP)’s storage on the cloud so we can access that as well as BigQuery which is just a cloud data warehouse so it’s very easy importing your data connecting to your different data sets within Google Cloud Platform (GCP)

Vertex AI Feature Store

the next is a vertex Artificial Intelligence (AI) feature store so the vertex Artificial Intelligence (AI) feature store becomes built in within Google Cloud Platform (GCP) and we can use it to serve and reuse features so the primary goal of a feature store is um if you’re doing complex transformations on your features you can go ahead and you can save those transformations and those transformed features in a feature store and then if you’re looking to build a different type of model using the same data or you’re looking to do some sort of similar transformations to different data sets or if you want to reuse the features for a different problem you can just go into the feature store and fetch those features and use it rather than trying to do all of those processes again it also helps from different teams re redoing feature engineering and they can just go in and use it from the feature store rather than doing that process again

Vertex AI Experiments

next is vertex Artificial Intelligence (AI) experiments so this makes tracking analyzing and discovering different run experiments very easy

Vertex AI Pipelines

vertex Artificial Intelligence (AI) pipelines so these are the primary this is the primary part that creates your Machine Learning (ML) workflow so these pipelines they also help to automate monitor and also govern your Machine Learning (ML) systems

AutoML

another feature within Vertex AI is AutoML so this feature can be used to create Machine Learning (ML) models without using any code so depending on the type of problem you have you specify like an objective or task that you have whether it’s a type of problem it is whether it’s classification multi-classification or a regression problem and then from there you basically define certain criteria like a metric that you want to minimize or maximize and then depending on those it goes ahead and builds a model up for you so it’s a great feature if you’ve never really used code before to build Machine Learning (ML) models it kind of takes care of everything on the back end and creates a model that best suited to your different parameters that you’ve provided

Vertex AI Vizler

vertex Artificial Intelligence (AI) vizier so this is a tool that helps to tune hyper parameters in Machine Learning (ML) models so currently sometimes if you’re working with the model in order to improve accuracy you might want to try different hyper different hyper parameters for the model so whether it’s decreasing the learning rate or if you want to add more estimators in a random forest model and many different things that things you can tune within a model so rather than doing a manual process or writing some sort of for loop that’s using different numbers you can use a you can use this which basically you define your scope of the parameters that you want to tune and you define the certain criteria in terms of what the different ranges are and then from there it basically just goes across those different selections and it trains many different models and then it gives you the best one um depending again on um a certain type of metric that you want to minimize or maximize whether that’s accuracy or your loss

Vertex AI Prediction

vertex Artificial Intelligence (AI) prediction so this is what actually deploys your Machine Learning (ML) model for online serving um this tool offer also offers a built-in tracking to track your model performance you can also go ahead and see how people are interacting with your model how many hits you’re getting requests or if there’s any errors when the request is sent for prediction

Vertex Explainable AI

vertex explainable Artificial Intelligence (AI) this is a really cool feature this really helps you understand the outputs of the model better it offers a more detailed evaluation metrics and uh you can also get feature attributions from this so it can tell you like how important each feature is for a certain prediction that is made

Key Takeaways Vertex AI

Unified ML platform, everything in one place

first is that it is a unified Machine Learning (ML) platform so everything is in one consolidated space so Google has brought together all of its uh Machine Learning (ML) and Artificial Intelligence (AI) APIs all together under one as Scott was showing that the list of the developer cheat sheet and it had all those different services under aiml it’s basically brought together all those apis and it’s uh it’s all under one place now before it was they were all different apis whereas now everything’s all connected and in one so this helps to build track and monitor the different stages than your data science life cycle and another big thing is that it can also help teams collaborate much easily so rather than some different teams and different people within teams building custom models or building models on their local machines you can build it in this cloud environment where everything is in one place so that way that if there’s any updates that are made to notebooks or models or pipelines that are saved in real time and they’re all saved under the specific projects that you specified for the

Build ML models without code (AutoML) and with code (custom models)

second thing is that models can be built without code and with code so as we’ve seen we showed an example of an AutoML model and then we showed the custom model that we had built uh we can see that there is high flexibility in which way you want to go so you can choose the rml version where you specify certain parameters what your task is like your target column in this case since we were using classification and it basically built up a model and then on the other hand if you want to build something that is more robust due to the control of like customization how you can really tune the model yourself control it what it’s outputting how it’s uh understanding your data that makes the model a little bit more flexible so depending on the complexity of the problem you have you can use AutoML it might be a good solution for it or you can also go the route of creating custom models and then also custom models can be created using Scikit-learn TensorFlow PyTorch all of the major frameworks and libraries that are used to build custom models

Can use the Vertex AI Workbench for ad-hoc model building

the next is the vertex Artificial Intelligence (AI) workbench for Ad-Hoc models so this is great to if you just wanted to try some sort of concept or some sort of just quick seeing if this this model or how a model works training up something quickly right Vertex AI workbench uses compute instances so you can also scale your development if you have some you have like you’re trying to build a large custom model it’ll also scale those instances and you can use GPUs along with the CPUs and it’ll really benefit in trading time rather than doing it locally

Create training pipelines using containerization to ensure versioning and consistency

another thing is creating training pipelines so uh using containerization so you can isolated environments can be built to ensure that versioning for packages and the consistency for the OS is present so this also eliminates like the dependencies that your model or your training code might have on certain environments and certain versioning like versioning and versioning of packages so that way that of your training pipelines they’re more versatile and you can basically run it on any instance so anytime you’re training you’re doing a training job on what vertex Artificial Intelligence (AI) does is it creates a compute instance and it’s like a virtual machine and it basically runs your code on that virtual machine so by incorporating and utilizing containerization you’re making sure that no matter which virtual machine that’s spinning up um so no matter what size or wherever whichever region it’s spinning it up on um you’re making you’re ensuring that your code will run on that virtual machine by making sure that you’re using a container and this is something that you have to do when you’re doing training you cannot you need to have either like a custom container or a container that it provides just to make sure that your code is able to run on different environments

Models can be easily logged, tracked, and deployed

then finally uh models can be logged and tracked and deployed very easily using vertex Artificial Intelligence (AI) so as we saw uh once uh the model was trained it was saved within our models dashboard we were able to basically deploy it to an endpoint and that is serverless so it’s not uh you’re not using a instance or your own or some sort of compute that’s being run um it’s all done from Google’s end so it’s a serverless endpoint uh it’s a serverless service and then you can create endpoints and to send like prediction requests and then get back the response as we saw um another feature is uh you can also monitor the model so you can capture model drift if there is any so model drift is um if there is data that is not representative of what you had trained your model on the meaning like things are changing or shifting your data is shifting towards something else and you’re seeing like a pattern so whatever data you might have trained your model on six months ago is not valid anymore and you need to retrain your model so this is a way of recognizing that is by monitoring the model and Google does that for you can monitor it and then you can identify like data drift and that way you can easily go ahead retrain and update the model to that same endpoint or you can create a whole new model without going through the whole process again since everything is already connected and it’s all unified.