Netflix, a company responsible for helping popularize paid TV and movie streaming, is conducting research into computer-simulated neural networks – machines designed to think in a way similar to the human mind.
According to a recent blog post from Netflix’s Alex Chen, Justin Basilico, and Xavier Amatriain, Netflix is finding ways to “train” large-scale neural networks. Contrary to the rigid logic of standard computers, neural networks simulate the learning processes that are thought to go on in the brain – requiring a neural network to be built, but also trained, to handle a certain workload.
A major customer of Amazon Web Services, Netflix has been experimenting in using the GPU power of public AWS cloud instances to train neural networks. This training involves developing machine learning algorithms by tuning the “hyperparameters” that drive the decision models. This is done by testing several different combinations of hyperparameters and picking the best one for the final model.
To more quickly train Netflix hopes that it can optimize this hyperparameter training using a method called Bayesian optimization. The company is experimenting with using the Spearmint package to perform Bayesian optimization across many GPUs using various tools to manage the hyperparameter workflow such as Celery, StarCluster, Jobman and HTCondor.
It’s currently unclear what application Netflix is hoping to building using this technology, but Netflix intends to share its experiences with neural networks to help grow the field.
The field of neural networks is still in an experimental stage and a set of common practices and community information sharing might be needed for this technology to take off.
Recently, a team of computer scientists at the University of Vienna outlined a system that provides on-demand neural network tools aimed at novice and experienced users, provided through a “Sky Compute” cloud made up of different services.
Neural networks are being used to perform tasks difficult with traditional computational approaches. Some applications include computer vision, speech recognition and artificial intelligence. With a surge in public and private research, neural networks could help solve new research puzzles, and also pave the way for new entrepreneurial products and services.