Initialization matters: Orthogonal Predictive State Recurrent Neural Networks

Authors: Krzysztof Choromanski, Carlton Downey, Byron Boots

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare the performance of OPSRNNs with that of LSTMs as well as conventional PSRNNs on a number of robotics tasks, and show that OPSRRNs are consistently superior on all tasks. Exhaustive experiments conducted on several robotics task confirm our theoretical findings.
Researcher Affiliation Collaboration Krzysztof Choromanski Google Brain kchoro@google.com Carlton Downey Carnegie Mellon University cmdowney@cs.cmu.edu Byron Boots Georgia Tech bboots@cc.gatech.edu
Pseudocode No The paper provides mathematical descriptions and textual explanations of methods, but does not include any formally structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about making the source code available or provide a link to a code repository.
Open Datasets Yes Swimmer We consider the 3-link simulated swimmer robot from the open-source package Open AI gym.2. Handwriting This is a digital database available on the UCI repository (Alpaydin & Alimoglu, 1998) Moving MNIST Pairs of MNIST digits bouncing around inside of a box according to ideal physics. http://www.cs.toronto.edu/ nitish/unsupervised_video/.
Dataset Splits Yes We collect 25 trajectories from a robot that is trained to swim forward (via the cross entropy with a linear policy), with a train/test split of 20/5. We use a train/test split of 40/8. We use 25 trajectories with a train/test split of 20/5. We use 1000 randomly selected videos, split evenly between train and test.
Hardware Specification No The paper mentions software used (Tensorflow framework in Python) but does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used for the experiments.
Software Dependencies No All models were implemented using the Tensorflow framework in Python. (No version numbers provided).
Experiment Setup Yes In two-stage regression we use history (similarly future) features consisting of the past (next) 2 observations concatenated together. We use a ridge-regression parameter of 10( 2) (this is consistent with the values suggested in Boots et al. (2013); Downey et al. (2017)). The kernel width is set to the median pairwise (Euclidean) distance between neighboring data points. We use a fixed learning rate of 0.1 for BPTT with a BPTT horizon of 20. We use a single layer PSRNN.