Transmitting model components via light waves: A deep learning technique to reduce latency in connected devices

Deep learning using light: Machine learning components encoded on light waves

It takes a few seconds for a smart device to answer if you ask it for a weather forecast. This latency is caused by the fact that connected devices do not have the memory or power necessary to run and store the massive machine-learning models required for the device understand what the user is asking. The model may be stored in a datacenter that is hundreds of miles from the device, and the answer is then computed there.

Researchers at MIT have developed a new way to compute directly on these devices. This method drastically reduces latency. The technique transfers the memory-intensive tasks of running a machine learning model to a server central where the components of the model can be encoded on light waves.

Fiber optics is used to transmit the waves to an attached device, allowing tons of data be sent in lightning speed through a network. The receiver uses a simple optical device to perform computations quickly using the model parts carried by these light waves.