Training Recurrent Neural Networks Online by Learning Explicit State Variables
Authors: Somjit Nath, Vincent Liu, Alan Chan, Xin Li, Adam White, Martha White
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We designed a sequence of experiments in real and synthetic problems to evaluate our new method compared with several common baselines and to highlight the robustness of our method to different truncation lengths, buffer sizes and number of updates. We investigate the empirical performance of our online FPP algorithm, particularly in terms of computation compared to truncated BPTT with varying truncation levels. |
| Researcher Affiliation | Academia | Department of Computing Science University of Alberta {somjit,vliu1,achan4,xzli,amw8,whitem}@ualberta.ca |
| Pseudocode | Yes | Algorithm 1 Fixed Point Propagation (FPP) |
| Open Source Code | No | The paper does not provide an explicit statement about releasing open-source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | The first problem is Sequential MNIST. The objective is to classify numerical digits based on a stream of pixel inputs. [...] Finally, we also include results on a a character prediction problem called Penn Tree Bank dataset. |
| Dataset Splits | No | The paper mentions evaluating performance over 'all training steps (online performance)' and provides total image/step counts for datasets (e.g., '1000 images' for MNIST, '5000 steps' for Penn Tree Bank), but does not specify explicit train/validation/test splits, percentages, or methodology for creating them. |
| Hardware Specification | No | The paper does not specify the hardware used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'RMSprop optimizer' but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | For all experiments, we use RMSprop optimizer and the learning rate is chosen over the set {0.0001, 0.0003, 0.001, 0.003, 0.01, 0.03} based on the average accuracy/loss. For all the algorithms, we used a constant buffer size of 100 and the trajectory length T for both T-BPTT(overlap and no overlap versions) and FPP. All algorithms use O(T) computation per step. We use a buffer size of 1000 and 10000 training steps. Network Type = simple RNN Hidden Units = 4 (for CycleWorld), and similar detailed specifications for other datasets in Appendix C. |