Regularizing Neural Networks with Meta-Learning Generative Models

Authors: Shin'ya Yamaguchi, Daiki Chijiwa, Sekitoshi Kanai, Atsutoshi Kumagai, Hisashi Kashima

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on six datasets showed that MGR is effective particularly when datasets are smaller and stably outperforms baselines.
Researcher Affiliation Collaboration Shin ya Yamaguchi (NTT), Daiki Chijiwa (NTT), Sekitoshi Kanai (NTT), Atsutoshi Kumagai (NTT), Hisashi Kashima (Kyoto University)
Pseudocode Yes Algorithm 1 Meta Generative Regularization
Open Source Code No No explicit statement or link providing access to the source code for the methodology described in this paper was found. Links provided (https://github.com/pytorch/vision, https://github.com/NVlabs/stylegan2-ada) are for third-party tools used in the research.
Open Datasets Yes We used six image datasets for classification tasks in various domains: Cars [12], Aircraft [27], Birds [28], DTD [29], Flowers [30], and Pets [31].
Dataset Splits Yes We randomly split a dataset into 9 : 1 and used the former as D and the latter as Dval.
Hardware Specification Yes We ran the experiments three times on a 24-core Intel Xeon CPU with an NVIDIA A100 GPU with 40GB VRAM
Software Dependencies No The paper mentions software like PyTorch, Kornia, and UMAP but does not provide specific version numbers for any of them. For example, 'We used the Image Net pre-trained weights of Res Net-18 distributed by Py Torch.'
Experiment Setup Yes We trained fθ by the Nesterov momentum SGD for 200 epochs with a momentum of 0.9, and an initial learning rate of 0.01; we decayed the learning rate by 0.1 at 60, 120, and 160 epochs. We trained Fϕ by the Adam optimizer for 200 epochs with a learning rate of 1.0 10 4. We used mini-batch sizes of 64 for D and 64 for Dp.