Neural Architecture Search

Neural architecture search (NAS) automates network design. But before automating, it helps to understand what you are designing and why each choice matters.

I combine multiobjective evolutionary algorithms with deep learning to search architectures that balance performance and cost, with applications in image super-resolution and text generation.

Explore it

Edit a neural network by hand — depth, width, activation, regularization, skip connections — and watch the diagram and the performance change. It's the manual process behind NAS, made visible: try a good, an average and a bad network, and feel the trade-offs.

In the optimization demo, in the NAS scenario, you'll see the other side: how a multiobjective evolutionary algorithm automates that search over the architecture space.

Why it matters: super-resolution

And why go to all this trouble to find good architectures? Here's the effect on a real task: image super-resolution. The same low-resolution input, reconstructed by a good architecture and by several bad ones —each with its characteristic artifact. This is what NAS aims to solve automatically.