Neural Architecture Search of SPD Manifold Networks
Authors: Rhea Sanjay Sukthanker, Zhiwu Huang, Suryansh Kumar, Erik Goron Endsjo, Yan Wu, Luc Van Gool
IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Statistical evaluation of our method on drone, action, and emotion recognition tasks mostly provides better results than the state-of-the-art SPD networks and traditional NAS algorithms. Empirical results show that our algorithm excels in discovering better performing SPD network design and provides models that are more than three times lighter than searched by the state-of-the-art NAS algorithms. ... Evaluation on three benchmark datasets shows that our searched SPD neural architectures can outperform handcrafted SPDNets [Huang and Van Gool, 2017; Brooks et al., 2019; Chakraborty et al., 2020] and the state-of-the-art NAS methods [Liu et al., 2018b; Chu et al., 2020]. |
| Researcher Affiliation | Academia | 1Computer Vision Lab, ETH Z urich, Switzerland 2PSI, KU Leuven, Belgium |
| Pseudocode | Yes | Algorithm 1 The proposed SPDNet NAS |
| Open Source Code | Yes | Source code link: https://github.com/rheasukthanker/spdnetnas. |
| Open Datasets | Yes | To keep the experimental evaluation consistent with the existing SPD networks [Huang and Van Gool, 2017; Brooks et al., 2019], we follow them to use RADAR [Chen et al., 2006], HDM05 [M uller et al., 2007], and AFEW [Dhall et al., 2014] datasets. |
| Dataset Splits | Yes | Following [Brooks et al., 2019], we assign 50%, 25%, and 25% of the dataset for training, validation, and test set, respectively. |
| Hardware Specification | No | The paper mentions "1 CPU day", "3 CPU days" and "8 GPU hours" as well as "AWS GPUs", but it does not specify the exact models or specifications of the CPUs or GPUs used for the experiments (e.g., Intel i7, NVIDIA V100). |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies, such as programming languages or libraries (e.g., Python 3.x, PyTorch 1.x). |
| Experiment Setup | Yes | All architectures are trained with a batch size of 30. Learning rate (η) for RADAR, HDM05, and AFEW dataset is set to 0.025, 0.025 and 0.05 respectively. |