STRODE: Stochastic Boundary Ordinary Differential Equation
Authors: Hengguan Huang, Hongfu Liu, Hao Wang, Chang Xiao, Ye Wang
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results show that our approach successfully infers event timings of time series data. Our method achieves competitive or superior performances compared to existing state-of-the-art methods for both synthetic and real-world datasets. Our experiments on CHi ME-5 show that our method outperforms ODE-RNN acoustic model in ASR. |
| Researcher Affiliation | Academia | 1National University of Singapore 2Rutgers University. |
| Pseudocode | No | The paper describes the model architecture and implementation details but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | We construct a synthetic video thumbnail dataset based on MNIST, the Rotating MNIST Thumbnail. ... CHi ME-5 was originally designed for the CHi ME 2018 challenge (Barker et al., 2018). |
| Dataset Splits | Yes | We use 5000 sequences for training 100 sequences for validation and 100 sequences for testing. ... For each subset, we generate 10000 video thumbnails for training, 1000 video thumbnails for validation and 1000 video thumbnails for testing. ... The Train, Dev and Eval include about 40 hours, 4 hours, and 5 hours of real conversational speech respectively. |
| Hardware Specification | Yes | In our experiments, the timing experiments use Py Torch package and are performed on Ubuntu 16.04 with a single Intel Xeon Silver 4214 CPU and a GTX 2080Ti GPU. |
| Software Dependencies | No | The paper mentions "Py Torch package" and "Ubuntu 16.04" but does not specify version numbers for PyTorch or other key software libraries used for reproducibility. |
| Experiment Setup | Yes | We repeat this training procedure across 3 different random seeds. ... using the Adam optimizer with a learning rate in the range [2 10 4, 6 10 4]. ... training all GMM-HMM. ... using BPTT (Werbos, 1990) and SGD with learning rates ranging from 0.13 to 0.19. We apply a dropout rate of 0.1 to the connections between neural network layers except that of ODE solvers. We reweight the importance of both KL terms and the prior log-likelihood. Both weights are set as 1 10 3. |