Researchers Make AI That Builds AI Much Faster, Cheaper
AI design usually guzzles computing power; this new method slashes the cost.
One of the biggest bottlenecks in AI isn't training the final model — it's designing the model itself. Neural Architecture Search (letting AI automatically figure out the best structure for another AI) normally involves testing thousands of candidate designs, each needing full training. That takes weeks on powerful supercomputers. This paper introduces a smarter shortcut: instead of fully training every candidate, the system uses a lightweight “surrogate” predictor that estimates how well a design would perform, based on early signs of learning ability.
To do this, the researchers created a new method called MFSPNet. It works like a swarm of bees exploring possible AI designs, using a clever prediction tool that watches how quickly the model improves in early training. This lets it rank designs reliably without the expensive full run. They also used a “block-based dense connection” strategy, which builds the AI from reusable building blocks that connect densely to earlier layers, helping information flow more smoothly — like a well-organized office where everyone talks to everyone.
On standard image-recognition tests, MFSPNet achieved accuracy similar to much more expensive methods — error rates as low as 1.91% on the SVHN digit dataset, and competitive results on larger, harder datasets like ImageNet. The key win: the entire architecture search took less than three GPU days, a fraction of the usual time and electricity. That's like finding a reliable new medicine in a week of lab work instead of a year.
The real-world impact is about democratizing AI. When designing a powerful AI costs less, smaller companies, researchers, and even hobbyists can afford it. It also reduces the environmental footprint of AI development. The authors note one limitation: their most complex test (ImageNet) ran only once due to computing limits, but even that single result strongly suggests the method scales well. Overall, this points to a future where AI builds smarter, faster, and more affordably — not just for tech giants, but for everyone.
- The new method finds good AI designs in under 3 days of computing — much faster than older techniques.
- It works by estimating performance early in training instead of running full tests, saving time and money.
- It hit accuracy similar to expensive methods on standard image tests, making smart AI more accessible to smaller teams.
Why It Matters
Cheaper AI design means faster innovation, lower costs, and fewer carbon emissions — benefits for every AI-powered product we use.