Learning Structured Representations with Hyperbolic Embeddings
Authors: Aditya Sinha, Siqi Zeng, Makoto Yamada, Han Zhao
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several large-scale vision benchmarks demonstrate the efficacy of Hyp Structure in reducing distortion and boosting generalization performance, especially under low-dimensional scenarios. |
| Researcher Affiliation | Academia | Aditya Sinha University of Illinois, Urbana-Champaign as146@illinois.edu Siqi Zeng University of Illinois, Urbana-Champaign siqi6@illinois.edu Makoto Yamada Okinawa Institute of Science and Technology makoto.yamada@oist.jp Han Zhao University of Illinois, Urbana-Champaign hanzhao@illinois.edu |
| Pseudocode | Yes | The pseudocode of Hyp Structure is shown in Algorithm 1 in Appendix B.2. |
| Open Source Code | Yes | The code is available at https://github.com/uiuctml/Hyp Structure. |
| Open Datasets | Yes | We consider three real-world vision datasets, namely CIFAR10, CIFAR100 [45] and Image Net100 [59] for training, which vary in scale, number of classes, and number of images per class. |
| Dataset Splits | Yes | We construct the Image Net100 dataset from this original dataset by sampling 100 classes, which results in 128,241 training images and 5000 validation images. |
| Hardware Specification | Yes | We implement our method in Py Torch 2.2.2 and run all experiments on a single NVIDIA Ge Force RTX-A6000 GPU. |
| Software Dependencies | Yes | We implement our method in Py Torch 2.2.2 and run all experiments on a single NVIDIA Ge Force RTX-A6000 GPU. |
| Experiment Setup | Yes | We follow the original hyperparameter settings from [39] for training the CIFAR10 and CIFAR100 models from scratch with a temperature τ = 0.1, feature dimension 512, and training for 500 epochs with an initial learning rate of 0.5 with cosine annealing, optimizing using SGD with momentum 0.9 and weight decay 10 4, and a batch size of 512 for all the experiments. |