ProxylessNAS
Designing a neural network for a phone or embedded board is not just about accuracy: the model must also respond quickly, fit in memory, and avoid draining the battery. ProxylessNAS is a method for automatically discovering neural-network architectures while accounting for the real device that will run them.
How it searches without a shortcut
Earlier neural architecture search methods commonly used a smaller “proxy” problem—such as lower-resolution images or fewer training epochs—to save search time. The winning design then had to be retrained and tested at full scale, where its apparent advantage could disappear. ProxylessNAS instead searches directly on the target task and target hardware, such as ImageNet classification on a mobile GPU or phone CPU.
Its key mechanism
ProxylessNAS builds a large supernetwork in which each layer has several candidate operations: for example, different mobile convolution blocks or kernel sizes. Activating every candidate at once would consume too much GPU memory. Its path binarization technique activates and trains only one candidate path per layer for a training step, while still learning which choices are best. The search objective combines:
- Task loss, which rewards accurate predictions.
- Measured or estimated latency, which penalizes architectures that run too slowly on the intended device.
Why this matters on edge hardware
A model with fewer operations is not automatically faster: a chip may handle one convolution shape efficiently and another poorly. ProxylessNAS selects architectures based on device behavior rather than abstract operation counts. This helps produce compact mobile networks that meet a latency budget—for example, keeping an offline smart-camera detector responsive or ensuring a phone vision feature does not raise heat enough to trigger throttling. The resulting architecture can then be deployed through runtimes such as TensorFlow Lite or ONNX Runtime, with its structure already chosen for practical hardware limits.
ProxylessNAS is a hardware-aware neural architecture search method that learns efficient network architectures directly for a target device, optimizing both accuracy and measured latency. It avoids using proxy networks or proxy tasks during search, reducing the gap between searched and deployed models. This matters for edge AI because it produces architectures that meet real mobile or embedded latency and resource budgets rather than performing well only in simulation.
Imagine designing a bicycle by testing it on the actual hills and streets where it will be ridden, rather than judging it only in a showroom. ProxylessNAS does something similar for AI models.
It helps create AI that is not just accurate, but also genuinely practical on a specific phone, camera, or small sensor device. Instead of choosing a model design based only on estimates, it considers how quickly the model runs on the real target hardware. That matters because a model that looks good on paper may be slow, drain the battery, or make a device feel unresponsive. ProxylessNAS aims to find designs that fit the device’s real-world limits.