Reproduce “Deep Signature Transforms” on OpenResearch

About this paper

Patric Bonnier, Patrick Kidger, Imanol Perez Arribas, Cristopher Salvi, Terry Lyons

Researchers from the University of Oxford and The Alan Turing Institute introduce Deep Signature Transforms, integrating the signature transform as a learnable, differentiable layer within deep neural networks, allowing for data-dependent augmentations and multi-layer architectures. This framework achieved substantially lower mean squared error in Hurst parameter estimation and successfully solved a non-Markovian reinforcement learning task where traditional recurrent networks failed.

Key insights

• By making the stream augmentation function learnable, the deep signature model can adaptively extract more relevant information from sequential data, overcoming the limitations of fixed pre-processing.
• The proposed deep signature model effectively combines the strong mathematical foundations of the signature transform (e.g., unique path characterization, universal nonlinearity) with the empirical success and flexibility of deep neural networks.
• The introduction of 'lifts' provides a mechanism to maintain the stream-like structure, allowing for the creation of deep, multi-layer architectures based on the signature transform for the first time.

Summary by alphaXiv.