Organizations using Amazon Web Services Inc.’s Amazon SageMaker platform to build artificial intelligence models now have access to a wealth of new features that will make machine learning projects simpler in several ways. 

The new features made their debut at AWS re:Invent today. Some of the capabilities are generally available, while others are in preview.

Accelerating AI projects

The SageMaker platform encompasses numerous services that not only provide features for developing neural networks, but also help with the other steps involved in AI projects. Those other steps include creating training datasets, one of the most time-consuming tasks in the entire AI development process. AWS is promising to speed up the task with a new service called Amazon SageMaker Ground Truth Plus.

Training datasets consist of records similar to the files an AI model will be expected to process once it’s running in production, plus associated labels. The labels provide useful contextual information. For example, if a retailer is building an AI to automatically sort merchandise by product category, the training dataset might consist of pre-sorted items with labels describing their product categories. 

Training datasets usually have to be created manually. As a result, the process can take a significant time and effort, especially in complex AI projects. With the new Ground Truth Plus service, AWS is promising to streamline the task by providing access to a network of experts who can add labels to a company’s dataset. Organizations only need to provide the data with which they plan to train their AI models and specify how labels should be added.

Ground Truth Plus features an AI system that learns from the experts who label a company’s data. Over time, the system can learn to complete some tasks automatically. “This means that there’s less need for a human to spend as much time creating each individual label for every object of interest in a dataset,” explained AWS Senior Developer Advocate Sean Tracey. “Less time spent on labeling means lower costs for you, and it also means a quicker turnaround in creating a dataset that can be used for training a model – all without sacrificing quality.”

In the SageMaker platform, software teams can carry out a large number of AI development and deployment tasks through a service called Amazon SageMaker Studio. As of today, the service lends itself to an even broader range of tasks thanks to a set of new updates.

Developers often run AI models they create in SageMaker Studio on another AWS service called Amazon EMR, which provides access to distributed data processing clusters. Developers can now create Amazon EMR clusters directly from the SageMaker Studio interface. They may configure the infrastructure manually, or use infrastructure templates supplied by their company’s information technology team, which is made possible by a new feature AWS also added today.

Troubleshooting technical issues that may arise during AI projects is now simpler as well. “We’ve built the ability to connect to, debug, and monitor Spark jobs running on an Amazon EMR cluster from within a SageMaker Studio Notebook,” Tracey wrote. Apache Spark is an open-source analytics engine widely used in AI projects.

For developers new to AWS and AI who wish to learn the ropes, the cloud giant is introducing a free service called Amazon SageMaker Studio Lab. It allows users to run AI models on the Amazon.com Inc. unit’s cloud using the popular Jupyter Notebook coding tool. 

Amazon SageMaker Studio Lab doesn’t require any cloud infrastructure management know-how or even an AWS account to use. Developers who apply for access and receive it from AWS simply have to specify whether they require an instance with a central processing unit or a graphics processing unit. The service offers a choice between 12 hours of CPU time or four hours of GPU time per user session.

Simplified AI training and deployment 

SageMaker users are receiving a new feature called SageMaker Training Compiler that promises to reduce the time it takes to train AI models by up to 50%. The feature boosts performance by optimizing AI models’ code to run more efficiently on GPUs.

In most machine learning projects, developers write their neural network using a programming language such as Python. They then leverage a framework such as TensorFlow to translate the Python code into low-level instructions that can be understood by GPUs. As part of the translation process, the code is tweaked to run faster, but AWS says that TensorFlow and similar frameworks only deliver limited performance improvements because they use general-purpose optimization methods. 

SageMaker Training Compiler takes a different approach. “It can take even the most skilled GPU programmers months to create custom kernels for each new model and optimize them. We built SageMaker Training Compiler to solve this problem. Today’s launch lets SageMaker Training Compiler automatically compile your Python training code and generate GPU kernels specifically for your model,” Tracey explained.

AWS is also making it easier to deploy AI models. A feature called SageMaker Serverless Inference allows developers to deploy neural networks for inference without having to configure or maintain any infrastructure for the task.

If developers do wish to set up the infrastructure manually, they can use another new capability called SageMaker Inference Recommender to ease the process. The feature suggests the most suitable instance types for an AI project. 

Image: AWS