Neural Architecture Search on the Gemini Enterprise Agent Platform explores very large spaces of possible model architectures to find the best combination of accuracy and constraints such as latency or memory.
What Is Neural Architecture Search?
With Agent Platform Neural Architecture Search you can search for optimal neural architectures in terms of accuracy, latency, memory, a combination of these, or a custom metric. The search space can be as large as 10^20 possible architecture choices.
According to the documentation, the underlying technique has successfully generated several state of the art computer vision models in the past years, including NASNet, MnasNet, EfficientNet, NAS-FPN, and SpineNet.
Google explicitly describes it as an experimentation tool: it is not a solution where bringing your data alone will produce good results.
Core Features
- Very large search spaces: Up to 10^20 architecture choices
- Multiple optimization targets: Accuracy with or without constraints such as latency, memory, or a custom metric
- Two-stage approach: Stage-1 search over proxy tasks, then full training of the top 10 models
- Customizable search spaces: Custom search spaces and rewards can be defined; per the documentation, customization can be done in approximately one to two days
- Prebuilt building blocks: Prebuilt vision search spaces and trainers, for example for MnasNet
Typical Use Cases
Architecture Search for Computer Vision
Use the prebuilt vision search spaces and trainers to optimize model architectures for image tasks.
Optimization Under Constraints
Search for architectures that meet latency or memory requirements alongside accuracy targets.
Custom, Non-Vision Search Spaces
Bring your own search spaces and trainers for tasks outside image processing.
Benefits
- Time savings on large search spaces: Per the documentation, when you already have a search space in mind, it can reduce at least approximately six months of engineering time exploring a large search space
- High customizability: Unlike supernet approaches (one-shot NAS or weight-sharing NAS), search spaces and rewards are largely free to define
- Control over the run: The search can be stopped manually at any point
- Transparent cost planning: The documentation describes example runs with trial counts, GPU types, quota requirements, and runtimes
Integration with innFactory
As a certified Google Cloud partner, innFactory supports you with Neural Architecture Search: assessing its fit for your use case, defining the search space, planning quotas, and evaluating trials.
Typical Use Cases
Technical Specifications
Frequently Asked Questions
What is Neural Architecture Search?
Per the documentation, Agent Platform Neural Architecture Search lets you search for optimal neural architectures in terms of accuracy, latency, memory, a combination of these, or a custom metric. The search space of possible neural architecture choices can be as large as 10^20.
Who is Neural Architecture Search for?
The documentation describes it as a high-end optimization tool and an experimentation tool, not a solution where you can just bring your data and expect a good result. It recommends having an in-house team for model tuning with a basic idea of the architecture parameters to modify, such as kernel size, number of channels, or connections. Google states it is meant for enterprise customers who can spend several thousand dollars on an experiment.
What is Neural Architecture Search not intended for?
Per the documentation, it is not for hyperparameter tuning such as learning rate or optimizer settings; it is only meant for architecture search, and the two should not be combined. It is not recommended with limited training data or for highly imbalanced datasets where some classes are very rare. Traditional methods such as hyperparameter tuning should be tried first.
How does a search run?
As the number of trials increases, the controller starts finding better models; reward variance and reward growth then decrease until convergence. The number of trials at convergence varies with search space size and is on the order of about 2000 trials. Each trial is designed as a smaller version of full training, called a proxy task, which runs for approximately one to two hours on two Nvidia V100 GPUs. After the search, the top 10 trials are selected and fully trained.
Is Neural Architecture Search limited to computer vision?
No. The documentation notes that Neural Architecture Search is not limited to vision use cases. Currently only vision-based prebuilt search spaces and prebuilt trainers are provided, but customers can bring their own non-vision search spaces and trainers.
Note: All product information on this page has been compiled with care, but is provided without guarantee and may be outdated or incomplete. Cloud services evolve rapidly — features, pricing, SLAs, and availability change frequently. Authoritative and up-to-date information can only be found on the official product page of Google Cloud (official documentation). This page does not represent an offer by Google Cloud.
