For more crisp and insightful business and economic news, subscribe to The Daily Upside newsletter. It's completely free and we guarantee you'll learn something new every day.
Knowledge is power. According to Amazon, that extends to machine learning models, too.
The company wants to patent a system that transfers the knowledge of one machine learning model to another. Essentially, Amazon's patent allows for the knowledge of one machine learning model that's completed its training to be transferred to another model that's undergoing training, thereby speeding up the training of AI models.
"Generally, training a machine learning model, such as a reinforcement learning model, from scratch requires a huge amount of time and computing resources," the filing notes. "Thus, it is desirable to have techniques to improve the learning speed of a machine learning model."
Amazon outlines two kinds of transfers: a representation transfer and an instance transfer. In a representation transfer, the knowledge that's given (such as how to perform a task or follow a policy) is based on the similarities between the "teacher model" and the "student model." Essentially, this system will distill the knowledge transferred to only the relevant information depending on those similarities.
Alternatively, in an instance transfer, the conveyed knowledge comes from "sampled trajectories" of the teacher model, or basically exact examples of its own inputs and outputs. The system automatically switches between doing representation transfers and instance transfers depending on the circumstance.
Amazon noted that this method could be applied to various AI training processes, including neural networks used for image processing and speech recognition.
Knowledge transfer is far from a new technique in the machine learning community, said Rijul Gupta, co-founder of AI communications company Deep Media. Gupta gave the example of an AI model that can play chess versus one that can play checkers: The games are somewhat similar, so transferring knowledge between them wouldn't be difficult.
Most of Amazon's patent is quite broad, Gupta said, so it's unlikely that it would get approved without a decent amount of pushback and narrowing. In its current form, granting it would "essentially cripple everyone else's machine learning development," Gupta said.
But the filing isn't entirely old news. One piece that's novel is its ability to automatically choose which teacher model is the best fit for training the student model without having to actually go through training, said Gupta. Typically, this process is a costly and tedious process that's done by hand.
Secondly, Amazon's patent discusses a way to bridge the gap between teacher and student models that are more distinct from each other. For example, rather than making the chess algorithm play checkers, Gupta said, it could "take an algorithm that's built to play chess and make it play Fortnite."
Using transfer techniques, he said, could speed up machine learning development by as much as 95%, helping it catch companies like Google, Microsoft and OpenAI. Meanwhile, securing the patent in a more narrow form could send the company's AI development soaring, making it "very difficult for any of Amazon's competitors to catch up to them."
Amazon has thrown its hat into the AI ring in more ways than one: The company pledged $100 million toward a AWS Generative AI Innovation Center, put together a new team focused on large language models, and has filed for multiple AI-related patents. As it stands, though, it remains a step behind the AI heavyweights.
That said, Amazon's real strength lies in AWS, said Gupta, with tools like SageMaker and Bedrock, as well as Inferentia and Trainium, its custom in-house chips for training AI. Plus, as AWS CEO Adam Selipsky told The Verge in a podcast last week, "Cloud and AI are not two different things. They're really just two of the many faces of the same thing."
"There's an old saying that when there's a gold rush, you should sell shovels," said Gupta. "And Amazon, right now, is selling shovels."
Have any comments, tips or suggestions? Drop us a line! Email at [email protected] or shoot us a DM on Twitter @patentdrop. If you want to get Patent Drop in your inbox, click here to subscribe.
