Improving Gradient Flow with Unrolled Highway Expectation Maximization

Authors: Chonghyuk Song, Eunseok Kim, Inwook Shim9704-9712

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We achieve significant improvement in performance on several semantic segmentation benchmarks and empirically show that HEMNet effectively alleviates gradient decay.
Researcher Affiliation Industry Ground Technology Research Institute, Agency for Defense Development {chonghyuk.song, a18700, iwshim}@add.re.kr
Pseudocode No The paper describes algorithmic steps using equations (e.g., E-step and M-step equations), but it does not present a structured pseudocode block or a clearly labeled algorithm.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets Yes We use Res Net (He et al. 2016a) pretrained on Image Net (Russakovsky et al. 2015) with multi-grid (Chen et al. 2017) as the backbone network. We set the step size to η = 0.5 for PASCAL VOC (Everingham et al. 2010) and PASCAL Context (Mottaghi et al. 2014), and η = 0.6 for COCO Stuff (Caesar, Uijlings, and Ferrari 2018).
Dataset Splits No The paper mentions evaluating on the PASCAL VOC val set, but does not specify the train/validation/test splits (e.g., percentages or counts) used for their experiments.
Hardware Specification No The paper does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, TensorFlow, specific libraries).
Experiment Setup Yes We use Res Net-50 with an output stride (OS) of 16 for all ablation studies and Resnet-101 with OS = 8 for comparisons with other state-of-the-art approaches. We set the temperature σ2 = C following (Vaswani et al. 2017), where C is the number of input channels to HEMNet, which is set to 512 by default. The step size is set to η = 0.5 for PASCAL VOC (Everingham et al. 2010) and PASCAL Context (Mottaghi et al. 2014), and η = 0.6 for COCO Stuff (Caesar, Uijlings, and Ferrari 2018). We set the training iteration number to Ttrain = 3 for all datasets. We use the moving average mechanism (Ioffe and Szegedy 2015; Li et al. 2019) to update the initial Gaussian bases µ(0). We adopt the mean Intersection-over-Union (m Io U) as the performance metric for all experiments across all datasets.