Self-Supervised Contrastive Pre-Training For Time Series via Time-Frequency Consistency

Authors: Xiang Zhang, Ziyuan Zhao, Theodoros Tsiligkaridis, Marinka Zitnik

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the new method on eight datasets, including electrodiagnostic testing, human activity recognition, mechanical fault detection, and physical status monitoring. Experiments against eight state-of-the-art methods show that TF-C outperforms baselines by 15.4% (F1 score) on average in one-toone settings (e.g., fine-tuning an EEG-pretrained model on EMG data) and by 8.4% (precision) in challenging one-to-many settings (e.g., fine-tuning an EEG-pretrained model for either hand-gesture recognition or mechanical fault prediction), reflecting the breadth of scenarios that arise in real-world applications.
Researcher Affiliation Collaboration Xiang Zhang Harvard University xiang_zhang@hms.harvard.edu Ziyuan Zhao Harvard University ziyuanzhao@college.harvard.edu Theodoros Tsiligkaridis MIT Lincoln Laboratory ttsili@ll.mit.edu Marinka Zitnik Harvard University marinka@hms.harvard.edu
Pseudocode No The paper does not contain any explicitly labeled pseudocode or algorithm blocks. It describes the methods and loss functions in prose and mathematical notation.
Open Source Code Yes The source code and datasets are available at https://github.com/mims-harvard/TFC-pretraining.
Open Datasets Yes The source code and datasets are available at https://github.com/mims-harvard/TFC-pretraining. ... (1) SLEEPEEG [61] has 371,055 univariate brainwaves (EEG; 100 Hz) collected from 197 individuals. ... (8) EMG [66] consists of 163 EMG samples with 3-class labels implying muscular diseases. Dataset labels are not used in pre-training. Further dataset statistics are in Appendix D and Table 3.
Dataset Splits Yes We randomly split the dataset into 80% for training, 10% for validation, and 10% for testing.
Hardware Specification Yes Computational resources: All experiments were conducted on NVIDIA A100 GPUs.
Software Dependencies Yes We use PyTorch [73] with Python 3.8.
Experiment Setup Yes All models are trained for 100 epochs with Adam optimizer (β1=0.9, β2=0.999), a weight decay of 1e−4, and a batch size of 128. For fine-tuning, we use a learning rate of 1e−3 and train for 20 epochs. The learning rate for pre-training is 1e−4. We randomly split the dataset into 80% for training, 10% for validation, and 10% for testing.