Artificial intelligence at MIT learned to train the neural network faster than ever

In an attempt to “democratize the AI” scientists at mit have found a way to use artificial intelligence to much more effective learning systems machine learning — that is, neural networks. They hope that a new algorithm, which allows to save time and money, will allow resource-constrained researchers and companies to automate the design of neural networks. In other words, reducing the time and cost they could make this technique more accessible AI.

A new field of artificial intelligence involves the use of algorithms for automatic design of neural networks that are more accurate and effective than those designed by human engineers. But this technology neuron-architectural search (neural architecture search NAS) is costly, in terms of processing power.

Most modern NAS algorithm, recently developed by Google to work on a bunch of GPUs, spent 48 000 GPU hours to create a single convolutional neural network used for image classification and detection tasks. Google has the ability to simultaneously run hundreds of GPUs and other specialized hardware in parallel, but is not available for many others.

The algorithm of the NAS, presented the Massachusetts Institute of technology, can directly teach specialized convolutional neural network (CNN) for the target hardware platforms — working with a massive dataset of images is just over 200 GPU-hours, which significantly expands the potential use of these types of algorithms.

According to scientists, resource-constrained researchers and companies could benefit from the algorithm in the form of saving time and costs. The overall goal is “the democratization of AI,” says study co-author song Han, associate Professor of electrical engineering and computer science Microsystems Technology Laboratories at MIT. “We want the experts on artificial intelligence and non-experts effectively designed the architecture of neural networks with a simple solution that quickly works on specific hardware”.

However, he adds that such NAS algorithms will never replace engineers-people. “The goal is to get rid of repetitive and tedious work associated with designing and improving the architecture of neural networks”.

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

Leave a Reply

Your email address will not be published. Required fields are marked *