Temporal Domain Generalization with Drift-Aware Dynamic Neural Networks
Authors: Guangji Bai, Chen Ling, Liang Zhao
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, extensive experiments on several real-world benchmarks with temporal drift demonstrate the proposed method s effectiveness and efficiency. In this section, we present the performance of DRAIN against other state-of-the-art approaches with both quantitative and qualitative analysis. |
| Researcher Affiliation | Academia | Guangji Bai , Chen Ling & Liang Zhao Department of Computer Science Emory University, Atlanta, GA, USA |
| Pseudocode | Yes | A.2 OVERALL GENERATION PROCESS We summarize the detailed forward propagation of DRAIN as below: a1 = 0, m1 = G0(z), z N(0, 1) a1 = Gη(ω1), ω1 Fξ(h1), (m1, h1) = fθ(m0, a0) a1 = Gη(ω1), ω1 Go(h1), (m1, h1) = fθ(m1, a1) a2 = Gη(ω2), ω2 = Φ(ω2, {ω1}), ω2 = Fξ(h1), (m2, h2) = fθ(m1, a1) a2 = Gη(ω2), ω2 = Φ(ω2, {ω1}), ω2 = Fξ(h1), (m2, h2) = fθ(m2, a2) at = Gη(ω2), ωt = Φ(ωt, {ωt τ:t 1}), ω2 = Fξ(h1), (mt, ht) = fθ(mt, at), |
| Open Source Code | Yes | 1Our open-source code is available at https://github.com/Bai The Best/DRAIN. |
| Open Datasets | Yes | Rotated MNIST: This is an adaptation of the popular MNIST digit dataset Deng (2012)... Appliances Energy Prediction: This dataset Candanedo et al. (2017) is used to create regression models of appliances energy use in a low energy building. |
| Dataset Splits | No | The paper specifies training and testing domains, for example, for "Rotated Moons": "Domains 0 to 8 (both inclusive) are our training domains, and domain 9 is for testing.", but does not explicitly mention a separate validation dataset split. |
| Hardware Specification | Yes | The experiments in this paper were performed on a 64-bit machine with 4-core Intel Xeon W-2123 @ 3.60GHz, 32GB memory and NVIDIA Quadro RTX 5000. |
| Software Dependencies | No | The paper mentions specific optimizers (Adam) and network components (LSTM, ReLU, Sigmoid) but does not provide version numbers for any specific software libraries or programming languages used (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | The learning rate is set to be 1e 4. (A.1.3) The number of layers in the LSTM is set to be 10, and the network architecture of gωt consists of 2 hidden layers, with a dimension of 50 each. (A.1.3, 2-Moons settings) |