Latent Matters: Learning Deep State-Space Models

Authors: Alexej Klushyn, Richard Kurle, Maximilian Soelch, Botond Cseke, Patrick van der Smagt

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our evaluation includes experiments on the image data of a moving pendulum [13] and on the reacher environment [26], where we use angle as well as high-dimensional RGB image data as observations.
Researcher Affiliation Collaboration 1Technical University of Munich 2Machine Learning Research Lab, Volkswagen Group, Munich 3Eötvös Loránd University 4AWS AI Labs
Pseudocode Yes Algorithm 1 REWO for deep state-space models
Open Source Code No The paper does not provide any statement or link indicating that source code for the described methodology is publicly available.
Open Datasets Yes The pendulum dataset was originally introduced in [13] and consists of 500 sequences with 15 images each... The reacher dataset consists of 2000 sequences with 30 time steps each. We use two versions in our experiments: (i) partially observed system states, i.e. the angles of the first and second joint; and (ii) RGB images of 64 64 pixels in size.
Dataset Splits No The paper describes the datasets and some aspects of evaluation (e.g., 'MSE of 500 predicted sequences') and refers to 'test ELBO' in tables, but it does not specify explicit train/validation/test split percentages or sample counts for the datasets used.
Hardware Specification No The paper does not specify any hardware details such as GPU models, CPU types, or memory used for the experiments.
Software Dependencies No The paper mentions general software concepts like 'RNNs' and 'LSTMs' but does not list any specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We provide a heuristic for the simple determination of D0 in App. A.1... Furthermore, Alg. 1 allows us to efficiently learn the parameters of the VHP (ψ0, φ0) and the transition model (ψ) by dividing the CO process into two phases: an initial and a main phase... For this purpose, we apply a special update scheme for λ, introduced and explained in depth in [16]: λ(i) = λ(i 1) exp h ν fλ λ(i 1), D(i)(θ, φ) D0; τ1, τ2 D(i)(θ, φ) D0 i . In this context, i denotes the iteration step of the optimisation process. The function fλ is defined as fλ(λ, δ; τ1, τ2) = 1 H(δ) tanh (τ1 (1/λ 1)) τ2 H (δ), where H is the Heaviside function, and τ1 as well as τ2 are slope parameters.