MIT artificial intelligence learned how to train neural networks faster than ever

In an attempt to "democratize AI" scientistsMassachusetts Institute of Technology found a way to use artificial intelligence for much more efficient learning of machine learning systems — that is, neural networks. They hope that the new algorithm, which will save time and money, will allow resource-limited researchers and companies to automate the design of neural networks. In other words, by reducing time and cost, they could make this AI technique more accessible.

Neural networks learn faster

A new field of artificial intelligence includesthe use of algorithms for the automatic design of neural networks that are more accurate and efficient than those developed by human engineers. But this neural architecture search technology (NAS) is expensive in terms of computing power.

Most recent NAS algorithm, recentlydeveloped by Google to work on a heap of graphics processors, spent 48,000 GPU-hours to create one convolutional neural network, which is used for image classification and detection tasks. Google has the ability to simultaneously run hundreds of graphics processors and other specialized equipment in parallel, but this is not available for many others.

NAS algorithm presented by Massachusettsas a technological institute, it can directly train specialized convolutional neural networks (CNN) for targeted hardware platforms — when working with a massive set of image data — in just 200 GPU-hours, which greatly expands the potential use of these types of algorithms.

According to scientists, resource limitedResearchers and companies could benefit from the algorithm in the form of time and cost savings. The overall goal is to “democratize AI,” says study co-author Song Khan, an assistant professor of electrical engineering and computer science at Microsystems Technology Laboratories at MIT. "We want both artificial intelligence experts and non-specialists to effectively design neural network architectures with a simple solution that quickly works on specific equipment."

However, he adds that such NAS algorithmsnever replace human engineers. "The goal is to get rid of repetitive and tedious work associated with the design and improvement of the architecture of neural networks."

Well, all this only accelerates the onset of the overallartificial intelligence. By the way, read our material about Demis Hassabis, the founder of DeepMind - one of the most promising companies in the field of AI.