University of Pennsylvania researchers have developed a new deep neural network processor. It is able to process, therefore to recognize and classify, billions of inputs per second. To do this, the chip harnesses the power of light: yes, it is a photonic processor. In name and in fact.
The study behind this marvel of optical engineering was recently published in Nature, by the researchers of the School of Engineering and Applied Sciences of the University of Pennsylvania. We link it to you here.
How does the photonic chip work?
This type of autonomous chip analyzes data in a similar way to that of our brain neurons, trained to recognize specific patterns of stimuli or input. This translates into the ability to detect the presence of objects, but above all to recognize and classify them: mechanisms like this are the basis of technologies we use every day, such as facial recognition and audio transcription, to name a few. To do this, the chip must have at least four characteristics, which are also what the researchers have tried to optimize:
- It must be able to convert optical signals into electrical signals;
- Likewise, it must be able to convert the input data into binary format;
- Must have large memory modules;
- It must have a linear calculation method.
Firooz Aflatouni, associate professor of electrical and systems engineering who carried out the study together with the graduate student Alexander J. Geers, explains: "Our chip processes information through what we call 'calculation by propagation '. In short, the calculations happen when the light travels through the chip ”.
The features provided allow the photonic chip to classify and recognize images almost instantaneously. The processing time of a single image? About half a nanosecond (a nanosecond is a billionth of a second). Imagine him at work recognizing the faces of a stadium audience. One by one. Before you can close your eyelids.
The study authors rate the photonic chip's performance as comparable to that of a high-end graphics processing unit (GPU) for image classification.
Yes, interesting ... But what is it for?
This photonic chip, which scientists say is the first to use light signals to develop such processing power, offers numerous advantages. First: lower energy consumption. According to: fewer bottlenecks than the deep neural networking technologies used today.
The chip still has limitations, especially related to its resolving power: for now it is able to work on images composed of a few pixels. The authors of the study already have in mind how to further increase the performance of the photonic chip, in particular by acting on the bandwidth of the micro ring modulators and photodioids in the optoelectronic layer.
With a bit of tricks and further implementations it will therefore be possible to further accelerate the photonic chip, and reach processing times of less than 100 picoseconds. Well… it doesn't end here!
