Our first product is a hardware co-processor called the Optical Processing Unit – or OPU -. It is designed to boost some of the most compute-intensive tasks in Machine Learning. The OPU can just be plugged onto a standard server or workstation, and accessed through a simple toolbox that is seamlessly integrated within familiar programming environments. Full-scale OPU prototypes are already available to selected users through the LightOn Cloud. We are opening our registration for those researchers and data scientists interested in trying out our technology on our cloud. The sign-up page is here.
Matrix-vector multiplications are amongst the most important elementary computing blocks in Machine Learning. For instance, Deep Learning schemes essentially stack such matrix-vector multiplications with non-linearities.
An OPU just does that: it multiplies the input data by a fixed matrix, passes through an element-wise non-linearity, and outputs the result. But because the OPU harnesses optics it can do this operation
- at massive data size
- very fast
- at minimal power consumption
What makes each OPU device literally unique is the fixed random matrix at the core if its computations, well fitted to the statistical learning of many Machine Learning / Artificial Intelligence schemes.
Examples of successful use cases of the OPU technology include:
- Image and Video Classification
- Recommender Systems
- Anomaly Detection
- Natural Language Processing
We regularly communicate our findings to the science community through preprints, presentations at conferences, blog posts and publications.
The proof of concept of our first generation prototype can be found at:
Random Projections through multiple optical scattering: Approximating kernels at the speed of light, Alaa Saade, Francesco Caltagirone, Igor Carron, Laurent Daudet, Angélique Drémeau, Sylvain Gigan, Florent Krzakala, https://arxiv.org/abs/1510.06664
Two different uses of the OPU for time-series analysis have been published in 2 papers :
Times series prediction with Echo State Networks
« Scaling up Echo-State Networks with multiple light scattering », Jonathan Dong, Sylvain Gigan, Florent Krzakala, Gilles Wainrib, IEEE Statistical Signal Processing Workshop (SSP), Freiburg, Germany, 2018, pp. 448-452, https://arxiv.org/abs/1609.05204
Online change-point detection in time series:
« NEWMA: a new method for scalable model-free online change-point detection », Nicolas Keriven, Damien Garreau, Iacopo Poli, https://arxiv.org/abs/1805.08061
Together with fellow scientists, we discuss this initial concept within the general framework of the links between machine learning and physics (including quantum computing)
Machine learning and the physical sciences https://arxiv.org/abs/1903.10563