LORD: Lower-Dimensional Embedding of Log-Signature in Neural Rough Differential Equations

Authors: JAEHOON LEE, Jinsung Jeon, Sheo yon Jhin, Jihyeon Hyeong, Jayoung Kim, Minju Jo, Kook Seungji, Noseong Park

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments with benchmark datasets, the improvement ratio by our method is up to 75% in terms of various classification and forecasting evaluation metrics.
Researcher Affiliation Academia Jaehoon Lee, Jinsung Jeon, Sheoyon Jhin, Jihyeon Hyeong, Jayoung Kim, Minju Jo, Seungji Kook, and Noseong Park Yonsei University Seoul, South Korea {jaehoonlee,jjsjjs0902,sheoyonj,jiji.hyeong,jayoung.kim, alflsowl12,2021321393,noseong}@yonsei.ac.kr
Pseudocode Yes Algorithm 1: How to train LORD-NRDE
Open Source Code Yes Our code is available in https://github.com/leejaehoon2016/LORD.
Open Datasets Yes We use six real-word dataset which all contain very long time-series samples. There are 3 classification datasets in the University of East Anglia (UEA) repository (Tan & Webb): Eigen Worms, Counter Movement Jump, and Self Regulation SCP2, and 3 forecasting datasets in Beth Israel Deaconess Medical Centre (BIDMC) which come from the TSR archive (Tan & Webb): BIDMCHR, BIDMCRR, BIDMCSp O2. We refer to Appendix B for detail of datasets.
Dataset Splits Yes Input: Training data Dtrain, Validating data Dval, Maximum iteration numbers max iter AE and max iter T ASK
Hardware Specification Yes Our software and hardware environments are as follows: UBUNTU 18.04 LTS, PYTHON 3.7.10, PYTORCH 1.8.1, CUDA 11.4, and NVIDIA Driver 470.42.01, i9 CPU, and NVIDIA RTX A6000.
Software Dependencies Yes Our software and hardware environments are as follows: UBUNTU 18.04 LTS, PYTHON 3.7.10, PYTORCH 1.8.1, CUDA 11.4, and NVIDIA Driver 470.42.01, i9 CPU, and NVIDIA RTX A6000.
Experiment Setup Yes The number of layers in the encoder, decoder and main NRDE, Ng, Nf, and No of Eqs. 9 to 11, are in {2, 3}. The hidden sizes, hg, hf, and ho of Eqs. 9 to 11, are in {32, 64, 128, 192}. The coefficients of the L2 regularizers in Eqs. 13 and 14 are in {1 10 5, 1 10 6}. The coefficient of the embedding regularizer, ce in Eq. 13 is in {0, 1, 10}. The max iteration numbers, max iter AE and max iter T ASK in Alg. 1, are in {400, 500, 1000, 1500, 2000}. The learning rate of the pre-training and main-training is 1 10 3.