Multi-level Residual Networks from Dynamical Systems View
Authors: Bo Chang, Lili Meng, Eldad Haber, Frederick Tung, David Begert
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply the proposed method to train Res Nets and Wide Res Nets for three image classification benchmarks, reducing training time by more than 40% with superior or on-par accuracy. |
| Researcher Affiliation | Collaboration | Bo Chang , Lili Meng & Eldad Haber University of British Columbia & Xtract Technologies Inc. Vancouver, Canada {bchang@stat, menglili@cs, haber@math}.ubc.ca Frederick Tung Simon Fraser University Burnaby, Canada ftung@sfu.ca David Begert Xtract Technologies Inc. Vancouver, Canada david@xtract.ai |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. |
| Open Datasets | Yes | Three widely used datasets are used for evaluation: CIFAR-10, CIFAR-100 (Krizhevsky & Hinton, 2009), and STL10 (Coates et al., 2011). |
| Dataset Splits | No | The paper provides sample counts for training and testing images but does not explicitly state the split information for a validation set, though it mentions "validation accuracy" in Figure 10's caption. |
| Hardware Specification | Yes | All the experiments are evaluated on machines with a single Nvidia Ge Force GTX 1080 GPU. |
| Software Dependencies | No | The networks are implemented using Tensor Flow library (Abadi et al., 2016). The paper mentions TensorFlow but does not provide a specific version number. |
| Experiment Setup | Yes | For our multi-level method, the models are interpolated at the 60th and 110th epochs. For baseline models, the learning rate cycle also restarts at epoch 60 and 110. The maximum and minimum learning rates ηmin and ηmax are set 0.001 and 0.5 respectively. For CIFAR-10 and CIFAR-100 experiments, the mini-batch size is 100. For STL-10 experiments, the mini-batch size is 32. We use a weight decay of 2 10 4, and momentum of 0.9. |