Google uses machine learning algorithms that will speed up the process of making artificial intelligence chips. The method used to produce these processors is faster than the manual process and can achieve the same results.
As Google shared in an interview with Nature, this project will be used for the first time in a commercial product. Specifically, in a future version of TPU (Tensor Processing Unit) chips. A component that can perform artificial intelligence tasks and will be used in Google’s data centers.
The task to be performed by machine learning algorithms is faster and more efficient than humans for a reason: The AI doesn’t have to learn or solve, but try a different combination. And so on until it gets it right.
The people responsible for making artificial intelligence chips have to study the processes, check how each component works, connect them together, use external tools, and so on. When a fault occurs, the engineering team looks for a way to fix it, which can take weeks to manufacture the component.
Making artificial intelligence chips is “like playing a game.”
In contrast, the process is much easier for AI, which takes the task “as a game,” Google engineers said. The mechanics are very simple: there is an integrated circuit to which small components must be attached. The goal is to find the right fit. If the artificial intelligence fails with that combination, it tries another. And so on until it gets the right one. The curious thing about AI is that it can form unimaginable combinations, which in turn help the component work properly.
This process helped Google train a reinforcement learning algorithm that was able to create a set of 10,000 designs of varying quality. Each design was given a “score,” and the algorithm uses that data to generate the correct blueprints for making the chips.
While it’s true that the results are not amazingly different from a human’s, the speed at which artificial intelligence can produce AI chips seems like a good reason to use this method.
Related posts:
