Recurrent Hidden Semi-Markov Model
Authors: Hanjun Dai, Bo Dai, Yan-Ming Zhang, Shuang Li, Le Song
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally justified our algorithm on the synthetic datasets and three real-world datasets, namely the segmentation tasks for human activity, fruit fly behavior and heart sound records. The R-HSMM with Viterbi exact inference significantly outperforms basic HSMM and its variants, demonstrating the generative model is indeed flexible. Moreover, the trained bi-RNN encoder also achieve similar state-of-the-art performances to the exact inference, but with 400 times faster inference speed, showing the proposed structured encoding function is able to mimic the exact inference efficiently. |
| Researcher Affiliation | Academia | Hanjun Dai1, Bo Dai1, Yan-Ming Zhang2, Shuang Li1, Le Song1 1 Georgia Institute of Technology {hanjundai, bodai, sli370}@gatech.edu, lsong@cc.gatech.edu 2 National Laboratory of Pattern Recognition, Chinese Academy of Sciences ymzhang@nlpr.ia.ac.cn |
| Pseudocode | Yes | Algorithm 1 Learning sequential VAE with stochastic distributional penalty method |
| Open Source Code | No | The paper does not provide any links to open-source code or state that code will be made available. |
| Open Datasets | Yes | Human activity This dataset which is collected by Reyes-Ortiz et al. (2016) consists of signals collected from waist-mounted smartphone with accelerometers and gyroscopes. Also: Drosophila Here we study the behavior patterns of drosophilas. The data was collected by Kain et al. (2013) with two dyes, two cameras and some optics to track each leg of a spontaneously behaving fruit fly. Also: Physionet The heart sound records, usually represented graphically by phonocardiogram (PCG), are key resources for pathology classification of patients. We collect data from Physio Net Challenge 2016 (Springer et al., 2015) |
| Dataset Splits | Yes | Without explicitly mentioned, we use leave-one-sequence-out protocol to evaluate the methods. Each time we test on one held-out sequence, and train on other sequences. Also: We do 5-fold cross validation for PN-Full. |
| Hardware Specification | Yes | We implemented with CUDA that parallelized for different RNNs, and conduct experiments on K-20 enabled cluster. |
| Software Dependencies | No | The paper mentions using 'Adam (Kingma & Ba, 2014)' for training and 'CUDA' for parallelization but does not provide specific version numbers for any software libraries or dependencies. |
| Experiment Setup | Yes | We set the truncation of max possible duration D to be 400 for all tasks. We also set the number of hidden states K to be the same as ground truth. Also: For the proposed R-HSMM, we use Adam (Kingma & Ba, 2014) to train the K generative RNN and bi-RNN encoder. To make the learning tractable for long sequences, we use back propagation through time (BPTT) with limited budget. We also tune the dimension of hidden vector in RNN, the L2-regularization weights and the stepsize. |