Google has updated its cloud AI platform, adding new features to AI Platform Prediction and AI Platform Training, enabling users to have more flexibility and efficiency with their artificial intelligence (AI) projects.

The AI platform brings together different tools designed to aid in machine learning development, from creating models to deployment. The two new updates announced are for AI Platform Prediction and AI Platform Training.

AI Platform Prediction enables application developers to utilise AI in a serverless environment, letting them choose from a set of Compute Engine machine types to run their model, and also add GPUs such as the inference-optimised, low latency NVIDIA T4.

Automated provisioning, scaling and serving

This means developers don’t have to understand machine learning frameworks or manage serving infrastructure in order to use the AI platform, as it handles the model’s provisioning, scaling and serving. Prior to this update, the Online Prediction required the user to choose from one of four vCPU machine types.

In addition, users can now log prediction requests and responses to BigQuery, using this to analyse and detect skews and outliers.

According to a blog post by Henry Tappen, product manager of Google’s Cloud AI Platform, this feature uses a new backend built on Google Kubernetes Engine, which aids in making the tool reliable, fast and flexible.

AI Platform Training

Google’s AI Platform Training has also received an update, now enabling data scientists to run a training script on their hardware without having to manage the underlying machines. To extend the types of models developers can train, Google has added custom containers and compute engine machine types for training, both of which are now generally available. 

Custom containers enable developers to launch any Docker container on the Google service, while Compute Engine allows developers to choose any combination of CPUs, RAM and accelerators.

Essentially, developers can now use existing AI models and training scripts with Google Cloud, without having to make code changes.