Amazon building ‘fake data’ to advance manufacturing

Big tech companies are entering a new era, racing to create fake – or synthetic – data to help further the advancement of machine learning.

Amazon executive Swami Sivasubramanian. Source: Supplied.

Data is one the most valuable assets in the world, but not all of it is real. Big tech companies are entering a new era, racing to create fake – or synthetic – data to help further the advancement of machine learning.

Amazon Web Services is one of the first in line to capitalise on the use of synthetic data, developing a new tool within its SageMaker cloud-based platform that is used by more than 100,000 companies to develop and deploy machine learning outcomes.

The new tool, available within Amazon SageMaker Ground Truth, allows a company to generate synthetic data of potential outcomes to help train artificial intelligence.

Its use cases are common in manufacturing companies which train AI and machine learning to develop their products, according to Swami Sivasubramanian, AWS vice president of analytics, database and machine learning.

A company that manufactures water bottles with the help of machine learning needs to train their AI with functioning products and defective products to create the best outcome, the executive said.

“Machine Learning needs to not just be training with an acceptable quality but also what a defect looks like. The tricky aspect of it is a defective product might be a one in 100,000 phenomenon but for you train a highly accurate model,” he said. “You need to feed in a variety of whatever the product is.”

Within the Amazon SageMaker platform, once a customer has provided sample images, AWS data scientists will customise those assets by artificially adding defects such as dents or scratches.

The process will turn what can be often be a several-month long process of collecting real data into just few days. “We can use 3D models and drawings for various combinations and that actually vastly save them immense amount of time,” Mr Sivasubramanian said.

The company also uses synthetic data to train its own robots, 520,000 of which pick and pack across Amazon warehouses.

ML is growing around the world including in Australia, where the form of AI has been taken up widely by customers including Swimming Australia and health insurer nib, said Amazon vice president of machine learning Brathin Saha.

Swimming Australia uses the technology to determine which athletes will perform best in relay races meanwhile nib uses ML and AI to further the comprehension of its chatbot nibby, he said.

“What is coming down the road tomorrow is all the ways in which machine learning is being used and our customers in Australia also gain something for that,” he said. “It’s here and now, it’s not the future.”

On Thursday (Friday AEDT), Amazon also announced the launch of a second new product called CodeWhisperer. The new product, which can generate code based on a developers previous code and ideas, was likened to a tool for bad coders by Mr Sivasubramanian.

“Simply it is an ML-powered code generation tool that directly listens to the idea and intent of the developers and using simple prompts like a upload these files to Amazon,” he said. “The purpose of the ML was to avoid bias by removing code recommendations that may be considered bias and unfair.”

Amazon is expecting the use of synthetic data and the ML code generator to grow phenomenally over the next few years. “Synthetic data is one of the next big waves of innovation as it’s unlocking the hurdles we face today,” Mr Sivasubramanian said.

“Machine learning is one of the most transformative technologies we will encounter in our generation but it won’t reach its full potential until we make these technologies accessible to organisations of all sizes.

“We are still in early days – it’s very much day one and we just have woken up and haven’t even had a cup of coffee yet.”

Joseph Lam was guest of Amazon Web Services at re: MARS in Las Vegas.