Bidirectional Learning for Time-series Models with Hidden Units
Authors: Takayuki Osogami, Hiroshi Kajino, Taro Sekiyama
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the effectiveness of bidirectional training and the hidden units in Dy BMs through numerical experiments using synthetic and real time-series. We will show that the Dy BM with hidden units can be trained more effectively with bidirectional training and reduces the predictive error by up to 90 %. |
| Researcher Affiliation | Industry | 1IBM Research Tokyo, Tokyo, Japan. |
| Pseudocode | Yes | Algorithm 1 Specific steps of bidirectional learning evaluated in experiments |
| Open Source Code | No | The paper does not provide concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper. |
| Open Datasets | Yes | The first dataset is the monthly sunspot number1, which we will refer to as Sunspot. (1https://datamarket.com/data/set/22t4/) The second dataset is the weekly retail gasoline and diesel prices2, which we will refer to as Price. (2https://www.eia.gov/dnav/pet/pet_pri_ gnd_a_epm0_pte_dpgal_w.htm) The third dataset is the NOAA Global Surface Temperature3, (3V4.00 of Air Temperature (air.mon.anom.nc) from https://www.esrl.noaa.gov/psd/data/gridded/ data.noaaglobaltemp.html) |
| Dataset Splits | No | The paper specifies training and test splits for the datasets (e.g., 'the first 67 % of each time-series is used for training, and the remaining 33 % is used for test.' for Sunspot and Price), but it does not explicitly mention a separate validation dataset split. |
| Hardware Specification | No | The paper mentions general hardware specifications ('workstations having 48-64 GB memory and 2.6-4.0 GHz CPU') but does not specify exact CPU or GPU models, or other detailed hardware components required for replication. |
| Software Dependencies | No | The paper mentions 'a Python implementation' and the use of 'Ada Grad', but does not provide specific version numbers for Python or any other libraries or software dependencies used in the experiments. |
| Experiment Setup | Yes | The bias b is initialized to zero, and the weight (U, V, W, Z) is initialized with i.i.d. normal random variable with mean 0 and standard deviation of 0.01... Throughout the experiments, we set the decay rate of the eligibility traces in (6) to zero... The delay d and the number of hidden units are varied for each dataset. ... For Sunspot, d = 30 for Price, d = 3, and d = 2 for Temperature. |