Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Scalable Signature Kernel Computations via Local Neumann Series Expansions
Authors: Matthew Tamayo-Rios, Alexander Schell, Rima Alaifari
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method, Power Sig, in terms of accuracy, memory usage, and runtime. Specifically, we compute the self-signature kernel of randomly drawn two-dimensional Brownian motion paths on [0, 1] at increasing sampling frequencies, using sample lengths ℓ= 2k + 1 for k ≥ 0, constrained only by GPU memory; comparisons are made against the state-of-the-art KSig library [19]. |
| Researcher Affiliation | Academia | Matthew Tamayo-Rios Seminar for Applied Mathematics & AI Center ETH Zurich, Switzerland Alexander Schell Department of Mathematics Technical University of Munich, Germany Rima Alaifari Department of Mathematics RWTH Aachen University, Germany |
| Pseudocode | Yes | Algorithm: Power Sig 1: procedure PROCESSDIAGONALS(X[0 .. cols 1], Y [0 .. rows 1], order) 2: Initialize s(0)[0] [1, 0, . . . , 0] 3: Initialize t(0)[0] [1, 0, . . . , 0] 4: for d 0 to rows + cols 3 do |
| Open Source Code | Yes | The codebase for our method, including all implementation details, is provided in the supplementary material and publicly available at: https://github.com/geekbeast/powersig. |
| Open Datasets | Yes | Figure 4 shows train and test fits (two-day rolling average) for kernel-ridge regression (MAPE) on the public bitcoin pricing dataset featured in Salvi et al. in [17]. (B) UEA Eigenworms classification and RFF/low-rank baselines. We benchmark Power Sig and KSig-PDE against linear/RBF kernel SVMs and the recent specialized RFF-based method RFSF-TRP from [20] on the standard Eigenworms dataset with input window lengths L {16, 32, . . . , 1024}. |
| Dataset Splits | No | Figure 4 shows train and test fits (two-day rolling average) for kernel-ridge regression (MAPE) on the public bitcoin pricing dataset... As shown in Figure 5, Power Sig (and KSig-PDE up to L = 128 before OOM) remains competitive and rises to 61.1% accuracy at L = 1024, whereas RFSF-TRP attains a slightly higher peak of 62.5% at L = 128 but exhausts memory for larger L... (The text mentions train/test splits and benchmarks on datasets but does not provide specific percentages, sample counts, or detailed splitting methodology within the main content.) |
| Hardware Specification | Yes | Experiments were run on an NVIDIA RTX 4090 GPU (24 GB). |
| Software Dependencies | No | While this setup directly supported Py Torch and Cu Py, special considerations were necessary for JAX, as its XLA backend does not natively allow resetting peak memory usage between runs... (The paper mentions PyTorch, CuPy, and JAX but does not provide specific version numbers for these software components in the provided text.) |
| Experiment Setup | Yes | Unless noted, Power Sig truncation order for the tile-center local series (24) is fixed at 7, although higher orders are equally feasible. |