Google is at it again with their latest patent for a revolutionary type of neural architecture search for mobile CNN models. In an effort to reduce the power and memory needed to run machine learning models, Google has come up with a new way to design neural networks that are more efficient without sacrificing accuracy.
Google's patent, US20230244904A1, is a continuation of U.S. Application Ser. No. 17/495,398, having a filing date of Oct. 6, 2021. It is based on and claims priority to U.S. Provisional Application No. 62/756,254 having a filing date of Nov. 6, 2018. The patent seeks to create a novel factorized hierarchical search space that will permit layer diversity throughout the network.
The term “neural architecture search” is used to refer to a process of finding the best neural network structure or “architecture” for a particular task. For example, if a computer is given the task of recognizing images of cats, it needs to determine which neural network structure is best suited for the task. This is usually done by running multiple experiments with different structures and evaluating their performance.
The goal of Google’s patent is to make neural architecture search easier and more effective, especially for mobile applications. Mobile devices often have limited power and memory resources, so it’s important for the neural networks used on them to be as efficient as possible. To achieve this, Google’s patent proposes a factorized hierarchical search space. This means that the search space is divided into smaller parts, each with its own characteristics. This allows for a greater range of layer diversity, which can lead to better performance.
For example, if one layer is designed for image recognition, another layer might be designed for natural language processing. This layer diversity can help the neural network to be more efficient and accurate, since it is better equipped to handle different types of input.
Google’s patent also proposes a method for evaluating the performance of different architectures. This is important for ensuring that the best architecture is found for a given task. With this method, the performance of different architectures can be compared on a variety of metrics, such as accuracy, latency, and memory usage.
As with any patent, there is no guarantee that the technology described in US20230244904A1 will make it to the market. That said, Google’s breakthrough neural architecture search has the potential to revolutionize mobile machine learning models. It could lead to more efficient and accurate models that use less power and memory, making them better suited for use on mobile devices. It could also help to make machine learning more accessible, since mobile devices are becoming increasingly popular.
Ultimately, Google’s patent is an exciting development in the field of machine learning. With its factorized hierarchical search space, it could enable the creation of more efficient and accurate neural networks for mobile applications. It remains to be seen, however, if this patent will come to market and how it might impact the world of machine learning.