At the 12th LightOn AI Meetup, we hosted Hashem Ghanem, a Ph.D. student at GIPSA-Lab, Grenoble, and IMB, Dijon, who presented his work on🕸️ Fast Graph Kernel with Optical Random Features ⚡ to appear at ICASSP 2021.
Hashem has been one of the first users of our technology on the LightOn Cloud, and we are particularly proud of his work using Optical Processing Units being accepted to a prestigious conference.
Graph structures are ubiquitous in the real world, and a problem of interest with applications in 🏦 banking, 🧬 biology, and 📣 marketing is graph classification: given a training set of labeled graphs, once presented with a new graph, predict to which class it belongs.
State-of-the-art methods for this problem include subgraph-based algorithms, graph convolutional networks, and graph kernels. This last approach is particularly effective, however, it incurs an exponential computational cost 📈, even after acceleration strategies have been applied.
Random features are an efficient approximation method for kernels, and we just happen to know a way to compute them at the speed of light ⚡🙃 with Optical Processing Units!
In the paper, they have the proper time complexities with O-notation but a plot is worth a thousand equations:
And the speedup does not come at a cost, the results on graph classification tasks look pretty good too 🚀
The code used for this work is publicly available. If you want to reproduce these results, expand them or try out your latest idea, you can register to the LightOn Cloud for a Free Trial or apply to the LightOn Cloud for Research Program!