How AI Automatically Builds Efficient Neural Networks for Limited Devices

Google's patent describes a method for automatically designing efficient neural networks by optimizing their internal structure, called 'computational cells,' to meet specific limits on resources like memory or processing speed.

What it covers

This patent details a computer-implemented method for automatically designing neural networks that are efficient for specific tasks and hardware. It starts by defining 'computational cells' (Claim 1), which are like small, reusable blueprints for parts of a neural network, composed of nodes (representing data) and edges (representing operations). For each operation, the system replaces it with a 'linear combination of candidate operations' (Claim 2), where each candidate operation has a 'mixing weight' controlled by 'computational cell hyper parameters.' The system then repeatedly adjusts these hyperparameters and weights to optimize how well the network performs (minimize 'validation loss function') while strictly adhering to 'computational resource constraints' (Claim 3, 9), such as limits on memory, processing power (float point operations), or how fast it can make predictions (inference speed). Finally, it uses these optimized settings to generate the complete neural network (Claim 1). For example, this could be used to create a highly efficient AI model for a smartwatch that needs to recognize speech without draining its small battery or exceeding its limited memory.

What it doesn't cover

• —Does not cover manual design of neural network architectures by human engineers.
• —Does not cover architecture search methods that do not use 'computational cells' defined as directed graphs of nodes and edges (Claim 1).
• —Does not cover optimization methods that do not involve replacing operations with a 'linear combination of candidate operations' (Claim 2).
• —Does not cover optimization without specific 'computational resource constraints' (Claim 1, 9) like memory, FLOPS, or inference speed.
• —Does not cover methods that do not iteratively adjust 'computational cell hyper parameters and computational cell weights' during the optimization process (Claim 3).

The clever bit

The novelty lies in automatically searching for optimal neural network architectures by representing network operations as flexible 'linear combinations' of candidate options. This allows for simultaneous optimization of these combinations and the network's overall structure (computational cells) under explicit, user-defined resource limits, systematically balancing performance with real-world hardware constraints.

Why it matters

This technology is crucial for deploying powerful AI models on devices with limited processing power, such as smartphones, smart home devices, or embedded systems. By automatically finding efficient network designs, it reduces the need for human experts to manually fine-tune models, saving time and resources. This allows AI to be integrated into more everyday products without requiring expensive, high-end hardware, making AI more pervasive and accessible.

Real-world examples

• 1.AI models for smartwatches
• 2.On-device speech recognition on smartphones
• 3.Edge AI for security cameras and drones
• 4.Machine learning in embedded systems for industrial control
• 5.Efficient AI for Internet of Things (IoT) devices