Neural Markov Controlled SDE: Stochastic Optimization for Continuous-Time Data

Authors: Sung Woo Park, Kyungjae Lee, Junseok Kwon

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS", "Table 1: Evaluation of reconstruction tasks on the Physio Net/Speech Commands datasets.", "Figure 2: Ablation study on the effectiveness of the proposed method.
Researcher Affiliation Academia Sung Woo Park1, Kyungjae Lee2, Junseok Kwon3 1,3School of Computer Science and Engineering, Chung-Ang University, Korea 1,2,3Artificial Intelligence Graduate School, Chung-Ang University, Korea
Pseudocode Yes Algorithm 1 Neural Markov CSDE-TP
Open Source Code No We used open-source codes provided by the authors for comparison.
Open Datasets Yes Datasets. For the evaluations, we used Physio Net, Speech Commands, Beijing Air-Quality, and S&P500 Stock Market datasets. Refer to Appendix A.5 for data statistics and prepossessing procedures.
Dataset Splits Yes We used a half of time-series as the training dataset and the remaining parts as the test dataset.", "80% were used as training dataset and the remaining parts as the test dataset.", "used first 80% of temporal states to train the model and the remaining parts are used for prediction task.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or specific computer specifications) used for running its experiments.
Software Dependencies No To estimate the gradient of the MBcond loss, we estimated numerical gradients with the auto-grad library in Pytorch (Paszke et al. (2019)).
Experiment Setup Yes For the running and terminal costs (l and Ψ, respectively), we used the l2 distance, i.e., l(s, x) = x ys 2 2 and Ψ(x) = x y T 2. In all experiments, γ is set to 0.95. The neural network structure for each agent control consists of 2-layers of fully-connected layers, where each module has 128 latent dimensions. The latent dimension of each Linear layer was set to 128 in all experiments except for the prediction task with the Air Quality dataset (= 64).