Optical Artificial Intelligence — Shining a Light on the Future of AI Hardware

Jordan Harrod
Dec 28, 2018 · 2 min read

This week’s paper can be found here: Training of photonics neural networks through in situ backpropagation

When discussing artificial intelligence, we typically work in the domain of software — code running on a computer, an algorithm running in the background of a website, or a model that gathers your data to predict what advertisement you should see next. It is true that much of the artificial intelligence currently integrated into our lives is software, but the need to develop hardware for artificial intelligence is becoming increasingly pressing.

AI models work as well as they do because they are trained on large amounts of data (among other things), however, the computational resources needed to train such an algorithm are expensive, both financially and energetically, and could not be implemented on things like Apple Watches and iPhones without causing heat damage to both the technology and the person using it. Additionally, using these models requires data sharing, which has resulted in many of the data privacy concerns prevalent today. Developing hardware that can bypass this issue by training an artificial intelligence model on small devices will be one of the next milestones of the AI timeline.

Many researchers have turned to light, or integrated optical methods, as a potential solution to this problem. Specifically, they want to develop optical artificial neural networks, because we can already use light to perform the matrix calculations to neural networks. The harder part is the actual training of the model. Up until now, we have been either training a model on a computer before translating it to its optical counterpart, or using complex optical calculations to train the model, which would be incredibly inefficient on the types of large AI models typically used in small devices.

However, researchers at Stanford may have found a solution to this inefficiency problem. Using layers of optical interference units, which can perform matrix calculations based on an input of light and output an intensity pattern, they train their optical hardware model by using the output pattern to figure out how accurate the model is. Once this “cost” is determined, they send a new pattern back through the hardware to adjust the OIUs, until the hardware model is as accurate as desired.


Originally published at www.jordanharrod.com.

Science + Communication + Policy

Read about new research in medical devices, AI, and science policy

Jordan Harrod

Written by

PhD Student at Harvard/MIT. Teaching you about AI via youtube.com/c/everydAI, Writing about Science + Policy on Medium.

Science + Communication + Policy

Read about new research in medical devices, AI, and science policy