Collapsed Amortized Variational Inference for Switching Nonlinear Dynamical Systems
Authors: Zhe Dong, Bryan Seybold, Kevin Murphy, Hung Bui
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we compare our method to various other methods that have been recently proposed for time series segmentation using latent variable models. Since it is hard to evaluate segmentation without labels, we use three synthetic datasets where we know the ground truth for quantitative evaluation but we also qualitatively evaluate the segmentation on a real world dataset.Table 1. Quantitative comparisons (in % σ) for segmentation on bouncing ball and reacher task. |
| Researcher Affiliation | Industry | 1Google AI, Mountain View, California, USA 2Vin AI Research, Hanoi, Vietnam. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and additional information: https://sites.google.com/view/cavi-snlds. |
| Open Datasets | Yes | In this section, we use a simple dataset from Johnson et al. (2016). In this section, we consider a dataset proposed in the Comp ILE paper (Kipf et al., 2019). Salsa Dancing from CMU Mo Cap data3 with footnote 3 pointing to http://mocap.cs.cmu.edu/. |
| Dataset Splits | No | The paper mentions evaluating on a 'held-out dataset of size 32' and using '29 of them as the training data, and hold out the other for evaluation' for the Salsa Dancing dataset. However, it does not provide specific percentages or counts for distinct training, validation, and test splits required for full reproducibility, nor does it refer to predefined splits with citations for all datasets. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'implementations of r SLDS, SVAE, and KVAE provided by the authors' but does not specify any software names with version numbers (e.g., Python version, specific library versions like PyTorch or TensorFlow). |
| Experiment Setup | Yes | Using entropy regularization and annealing to encourage discrete state transitions. To encourage the model to utilize multiple states, we add an additional regularizing term to the ELBO that penalizes the KL divergence between the state posterior at each time step and a uniform prior... L(θ, φ) = LELBO(θ, φ) βLCE(θ, φ), (11) where β > 0 is a scaling factor. To further smooth the optimization problem, we apply temperature annealing to the discrete state transitions, as follows: p(st = k|st 1 = j, xt 1) = S( p(st=k|st 1=j,xt 1) τ ), where τ is the temperature. At the beginning stage of training, β, τ are set to large values. ... Over time, we reduce the regularizers to 0 and temperature to 1, according to a fixed annealing schedule. |