Structured Recurrent Temporal Restricted Boltzmann Machines

Authors: Roni Mittelman, Benjamin Kuipers, Silvio Savarese, Honglak Lee

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results using synthetic and real datasets demonstrate that the SRTRBM can significantly improve the prediction performance of the RTRBM, particularly when the number of visible units is large and the size of the training set is small. It also reveals the dependency structures underlying our benchmark datasets.
Researcher Affiliation Academia Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI Computer Science Department, Stanford University, Stanford, CA
Pseudocode Yes Algorithm 1 Contrastive Divergence training iteration for the spike-and-slab SRTRBM.
Open Source Code No The paper mentions using code from a prior work ("using the code that is available online and was described in Sutskever et al. (2008)") but does not provide a link or statement about making their own implementation code available.
Open Datasets Yes In this experiment, we used the CMU motion capture dataset1, which contains the joint angles for different motion types. 1http://mocap.cs.cmu.edu/ In this experiment, we used a dataset of historical weather records available at the National Climatic Data Center2. 2ftp://ftp.ncdc.noaa.gov/pub/data/ushcn/daily
Dataset Splits No The paper specifies training and testing splits for its datasets (e.g., "4000 videos... for training... 200 such videos for testing"), but does not explicitly mention a validation set or a three-way split including validation.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models, memory, or cluster specifications.
Software Dependencies No The paper does not provide specific version numbers for any software components, libraries, or programming languages used in the experiments.
Experiment Setup Yes The number of CD iterations during training was set to 25. We used a fixed learning rate of 10 3 with 10 CD iterations for training the Gaussian RTRBM, and a single CD iteration for training the spike-and-slab-based methods. The spike-and-slab hyperparameter α was set to 1. The SRTRBM regularization parameter β was set to 5 10 3 when assigning each observation to an independent node, and 10 2 when assigning observations to groups based on the body parts. We used a learning rate of 5 10 5, and SRTRBM regularization parameter β = 10 2. We used the spike-and-slab parameter α = 8.5 and α = 1 for the non-structured and structured versions of the RTRBM respectively.