At the Google I/O event, the company has announced a revamped cloud-based machine learning platform branded as Vertex AI.
It’s unusual for Google to announce cloud-related services at Google I/O, a forum for launching consumer and developer technologies. Since Cloud Next – the flagship cloud user conference – is postponed to October, Vertex AI found its place in I/O announcements.
This is the third iteration of the Google Cloud ML platform since its original launch. What prompted Google to launch Vertex AI? Let’s find out.
What is Vertex AI?
Vertex AI brings multiple AI-related managed services under one umbrella. Google Cloud has two different AI services – AutoML and custom model management.
The first is a set of AutoML tools for training models based on vision, text, and tabular datasets. Auto ML Vision, AutoML Natural Language, AutoML Tables are some of the services in this category.
Google Cloud AI Platform, the managed ML Platform as a Service (PaaS), enabled customers to use Jupyter Notebooks and a custom Python SDK to train custom models. The platform is based on the GCE infrastructure with AI accelerators – GPU and TPU. Some components of the ML PaaS are based on Kubeflow, a cloud native, open source project that brings the machine learning environment to Kubernetes.
MORE FOR YOU
Vertex AI combines the AutoML services and AI Platform to deliver a unified platform covering almost every aspect of MLOps – the operational framework for managing machine learning projects.
Apart from the unification of AutoML and Cloud AI Platform, Vertex AI adds brand new features, including Edge Manager, Feature Store, Model Monitoring, and Vizier. The new capabilities plug critical gaps that existed in Google’s AI portfolio.
Why did Google launch Vertex AI?
Despite owning the most successful machine learning and deep learning framework, TensorFlow, Google struggled to deliver data scientists and developers the right experience. It missed some of the core capabilities needed by enterprises to deploy and manage ML models in production.
The very first offering of an ML platform from Google was the Cloud ML Engine. It didn’t have a user experience and native integration with IDEs such as Jupyter Notebooks. Developers had to use the SDK or the command-line tool to submit jobs to ML Engine.
In 2019, Google launched a broad set of tools that augmented ML Engine under the brand of Cloud AI Platform. The most interesting part of the Cloud AI Platform lies in its tight integration with Kubeflow, a popular open source, cloud native machine learning project backed by the industry bigwigs.
But these investments didn’t help Google close the gap with competitive offerings – Amazon SageMaker and Azure ML.
AWS got it right the first time with Amazon SageMaker. It has been steadily adding features like SageMaker Studio, Autopilot, JumpStart, Feature Store, Model Monitoring, and SageMaker Neo. SageMaker has emerged as one of the most comprehensive ML platforms in the public cloud in just three years.
Like Google, Microsoft has gone back and forth with its ML platform strategy. But the most recent avatar of Azure ML is stable and appealing to the data science community. The visual designer for building no-code/low-code deep learning models, integrated AutoML, tight integration with Azure Synapse and Azure IoT Edge make it one of the robust enterprise ML platforms.
Google has been working hard to enhance its AI Platform. With Vertex AI, it inches closer to Amazon SageMaker and AzureML. Some of the capabilities like Feature Store, Model Management, and Vizier will help Google tell a better story to enterprises.
Along with Cloud AI Platform, Google has also launched AI Hub – a comprehensive collection of Jupyter Notebooks, reusable pipeline artifacts, models, and modules that could be deployed on both Kubeflow platform and Cloud AI Platform.
Soon after the announcement, the AI Hub product page and documentation became unavailable. It may resurface with new branding and assets targeting Vertex AI.
It also makes it unclear whether existing Kubeflow Pipelines definitions work with Vertex AI.
The integration and compatibility of Cloud AI Platform with Kubeflow enabled Google to extend the ML PaaS to Anthos to bring hybrid AI capabilities to enterprises. If Vertex AI Pipelines are not interoperable with Kubeflow Pipelines, Google loses the advantage of hybrid AI spanning on-premises, multi-cloud, and hybrid cloud environments.
Google hasn’t shared if parts of Vertex AI will be available on Anthos hybrid and multi-cloud platform.
At least in the initial release, Vertex AI is not integrated with Google Kubernetes Engine for distributed training and at-scale inference. With recent additions such as multi-instance GPU clusters, GKE becomes the ideal target for training and inference.
The addition of capabilities such as Feature Store and Model Monitoring is exciting for Google Cloud users. Feature Store acts as a repository for features extracted from real-time data streams and historical datasets residing in data lakes and data warehouses. Model Monitoring feature will help customers detect drift in model accuracy and automatically triggering the training jobs.
With Vertex AI, Google not only unifies its AI offerings but also closes the gap with similar offerings from its key competitors.