Neural Networks Have Gone to Plaid!

Artificial intelligence algorithms have had a lot of wins in recent years as they have become adept at tasks ranging from image classification to predictive text and medical diagnoses. A major factor that is slowing an even more rapid advancement of these algorithms is the technology on which they operate. Modern computing technology may be highly advanced and incredibly fast at crunching numbers, but it is important to remember that computers were not designed with running artificial neural networks, or any other machine learning algorithms, in mind. In fact, when you really examine all that a computer needs to do to run a neural network, it can look a bit like hammering in a nail with a screwdriver.

A group of engineers from the University of Pennsylvania thought through an image classification algorithm, and realized that there are four primary areas that unnecessarily consume time due to architectural constraints of computing systems. First, optical signals need to be converted to electrical signals, then those signals must be converted to binary data. This data must be stored in a large memory module, and computations then proceed in synchronization with clock pulses. To eliminate these slowdowns, they came up with a radically different approach to classifying images. And that approach made it possible to classify images at the speed of light. That equates to nearly two billion image classifications per second.

That result is incredibly impressive, but may also lead one to question why such speed is necessary. For a basic image classification task, it may not be — but problems involving classifying moving objects, 3D object identification, or recognizing microscopic cells can push beyond the limits of current computing technology. In cases such as these, new methods are essential. And based on past experiences, it is probably safe to say that there will always be a new class of problems that are just beyond the reach of current technological limits.

The researchers accomplished their feat by designing an optical deep neural network on a 9.3 square millimeter chip. This chip directly processes the light that bounces off the object of interest and provides a near instantaneous classification. Similar to the artificial neural networks we are familiar with, the optical network is constructed of many optical neurons. These neurons are connected by optical wires called “waveguides” to form a network. Multiple layers such as this are stacked on top of one another, and as light passes through each layer of the optical network, it works to further classify the light signal it is receiving.

As a proof of concept, the team designed an optical deep neural network chip capable of recognizing handwritten characters. Within half of one nanosecond, the chip was found to be capable of classifying characters with 93.8% and 89.8% accuracies when discriminating between two and four possible characters, respectively. It would be interesting to see a new chip built that can discriminate between a much larger character set, however, as the prototype has little real world utility until that can happen.

In addition to the speed offered by this approach, it also helps with privacy-related concerns because there is no need to store processed data, such as images. The team also notes that the method has applications beyond visual classifications — by converting, for example, sound or speech to optical signals, it could be fed into an optical neural network. But for now they are focused on 3D object classification, which is currently a bit out of reach for traditional methods.