Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Restricted Random Pruning at Initialization for High Compression Range
Authors: Hikari Otsuka, Yasuyuki Okoshi, Ángel López García-Arias, Kazushi Kawamura, Thiem Van Chu, Daichi Fujiki, Masato Motomura
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on image classification using CIFAR-10 and CIFAR-100 and node classification using OGBN-Ar Xiv show that Mi CA enhances the compression ratio and accuracy trade-offs compared to existing Pa I methods. |
| Researcher Affiliation | Academia | Hikari Otsuka EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology Yasuyuki Okoshi EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology Ángel López García-Arias EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology Kazushi Kawamura EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology Thiem Van Chu EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology Daichi Fujiki EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology Masato Motomura EMAIL Department of Information and Communications Engineering Tokyo Institute of Technology |
| Pseudocode | Yes | Algorithm 1 Number of Output Nodes Analysis Algorithm 2 Step 1 of Minimum Connection Construction at l-th layer Algorithm 3 Step 2 of Minimum Connection Construction at l-th layer Algorithm 4 Minimum Connection Assurance |
| Open Source Code | No | The paper mentions that 'each implementation is based on the code provided by Tanaka et al. (2020)' for image classification architectures and 'implement each architecture based on the codes provided by Wang et al. (2019); Huang et al. (2022)' for node classification architectures. However, there is no explicit statement or link provided for the open-source code of the proposed Mi CA methodology itself. |
| Open Datasets | Yes | We employ the CIFAR-10, CIFAR-100, and Image Net datasets in image classification. For CIFAR-10 and CIFAR-100, (Krizhevsky et al., 2009)... Russakovsky et al. (2015) and on node classification using OGBN-Ar Xiv (Hu et al., 2020). |
| Dataset Splits | Yes | For CIFAR-10 and CIFAR-100, 40,000 images are used as training data and 10,000 as validation data, while we use the default set split for Image Net. ... on the OGBN-Ar Xiv dataset split by default proportion. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or other computer specifications used for running the experiments. It only describes training parameters and methods. |
| Software Dependencies | No | The paper mentions using 'stochastic gradient descent (SGD) applying Nesterov s acceleration method (Nesterov, 1983) with a momentum of 0.9' and 'Adam (Kingma & Ba, 2014) and cosine learning rate decay'. However, it does not specify version numbers for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | CIFAR-10 and CIFAR-100 experiments are run five times with a batch size 128 for 160 epochs, and the Image Net experiment is run once with a batch size 256 for 90 epochs. For VGG-16, the weight decay is set to 0.0001, and the learning rate is started at 0.1 and multiplied by 0.1 after 60 and 120 epochs. For Res Net-20, the weight decay is set to 0.0005, and the learning rate is started at 0.01 and multiplied by 0.2 after 60 and 120 epochs. For Res Net-50, the weight decay is set to 0.0001, and the learning rate is started at 0.1 and multiplied by 0.1 after 30, 60, and 80 epochs. ... Each architecture is set to an initial learning rate of 0.001 and is trained for 400 epochs. Also, these experiments are run five times. |