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; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Spectral-Aware Reservoir Computing for Fast and Accurate Time Series Classification

Authors: Shikang Liu, Chuyang Wei, Xiren Zhou, Huanhuan Chen

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate the SARC on public benchmarks. As aforementioned, the Freq Res is not confined to a specific RC model. By default, we implement Freq Res based on a Bidirectional ESN (Bi ESN), which is demonstrated to be optimal across four conventional RC models in our ablation studies (Section 5.3)4. All experiments are conducted using Python 3.11 on a desktop with an Intel Core i7-14700KF CPU, and an NVIDIA Ge Force RTX 4090D GPU.
Researcher Affiliation Academia 1School of Computer Science and Technology, University of Science and Technology of China, Hefei, Anhui, China. Correspondence to: Xiren Zhou <EMAIL>, Huanhuan Chen <EMAIL>.
Pseudocode Yes Algorithm 1 Spectral-Aware Reservoir Computing
Open Source Code Yes Summarily, our main contributions are as follows2: Code is available at https://github.com/ZOF-pt/SARC.
Open Datasets Yes We use the full UCR Time Series Archive (Dau et al., 2019) with 128 datasets spanning various applications such as activity recognition, health monitoring, and spectrum analysis.
Dataset Splits Yes Key hyperparameters are determined through a five-fold cross-validation on the training set, selecting input scaling from {0.5, 1, 2, 4}, spectral radii from {0.4, 0.6, 0.8}, regularization ζ from {0.5, 1}, and leaky rates ranging from 0 to 0.8 in 0.2 increments. The reservoir size is set to 10, the connectivity is 1, and the threshold κ is set to 100. For classification, we concatenate the derived dynamic features with the max-pooled hidden states and feed them to a default Ridge classifier.
Hardware Specification Yes All experiments are conducted using Python 3.11 on a desktop with an Intel Core i7-14700KF CPU, and an NVIDIA Ge Force RTX 4090D GPU.
Software Dependencies Yes All experiments are conducted using Python 3.11 on a desktop with an Intel Core i7-14700KF CPU, and an NVIDIA Ge Force RTX 4090D GPU.
Experiment Setup Yes Key hyperparameters are determined through a five-fold cross-validation on the training set, selecting input scaling from {0.5, 1, 2, 4}, spectral radii from {0.4, 0.6, 0.8}, regularization ζ from {0.5, 1}, and leaky rates ranging from 0 to 0.8 in 0.2 increments. The reservoir size is set to 10, the connectivity is 1, and the threshold κ is set to 100. For classification, we concatenate the derived dynamic features with the max-pooled hidden states and feed them to a default Ridge classifier.