Heteroscedastic Temporal Variational Autoencoder For Irregularly Sampled Time Series

Authors: Satya Narayan Shukla, Benjamin Marlin

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed architecture on both synthetic and real data sets. Our approach outperforms a variety of baseline models and recent approaches in terms of log likelihood, which is our primary metric of interest in the case of probabilistic interpolation. Finally, we perform ablation testing of different components of the architecture to assess their impact on interpolation performance.
Researcher Affiliation Academia Satya Narayan Shukla & Benjamin M. Marlin College of Information and Computer Sciences University of Massachusetts Amherst Amherst, MA 01003, USA {snshukla,marlin}@cs.umass.edu
Pseudocode No The paper describes the model components and their mathematical formulations but does not provide any structured pseudocode or algorithm blocks.
Open Source Code Yes Implementation available at https://github.com/reml-lab/hetvae and The source code for reproducing the results in this paper is available at https://github.com/ reml-lab/hetvae.
Open Datasets Yes Physio Net Challenge 2012 (Silva et al., 2012) and MIMIC-III (Johnson et al., 2016) consist of multivariate, sparse and irregularly sampled time series data. We also perform experiments on the Climate dataset (Menne et al., 2016), consisting of multi-rate time series. ... Physio Net is freely available for research use and can be downloaded from https://physionet.org/content/challenge-2012/. MIMIC-III is available through a permissive data use agreement which can be requested at https://mimic.mit.edu/iii/gettingstarted/. ... The dataset is available for download at https://cdiac.ess-dive.lbl.gov/ftp/ushcn_daily/. ... The dataset is available for download at https://archive.ics.uci.edu/ml/datasets/ individual+household+electric+power+consumption.
Dataset Splits Yes We randomly divide the real data sets into a training set containing 80% of the instances, and a test set containing the remaining 20% of instances. We use 20% of the training data for validation.
Hardware Specification Yes All experiments were run on a Nvidia Titan X and 1080 Ti GPUs.
Software Dependencies No The paper mentions using Adam Optimizer and Nvidia GPUs but does not specify versions for programming languages, deep learning frameworks (e.g., Python, PyTorch), or other libraries.
Experiment Setup Yes We fix the time embedding dimension to de = 128. The number of embeddings H is searched over the range {1, 2, 4}. We search the number of reference points K over the range {4, 8, 16, 32}, the latent dimension over the range {8, 16, 32, 64, 128}, the output dimension of Un TAND J over the range {16, 32, 64, 128}, and the width of the two-layer fully connected layers over {128, 256, 512}. In augmented learning objective, we search for λ over the range {1.0, 5.0, 10.0}. We use the Adam Optimizer for training the models. Experiments are run for 2, 000 iterations with a learning rate of 0.0001 and a batch size of 128.