Csaba Toth
Randomized signature features for scalable approximation of the signature kernel
Abstract
We revisit signature kernel algorithms and provide a randomized construction that 1) converges to the true kernel in probability as the embedding dimension increases, 2) give scalable algorithms, 3) shows little loss of performance or even improvements compared to all other variants while exhibiting much better computational complexities. The key idea is the use of random Fourier features and structured random projections for tensors.
Our speaker
Csaba Toth is a postgraduate researcher at the University of Oxford.
