Hierarchically Clustered Representation Learning
Authors: Su-Jin Shin, Kyungwoo Song, Il-Chul Moon5776-5783
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted evaluations with image and text domains, and our quantitative analyses showed competent likelihoods and the best accuracies compared with the baselines. |
| Researcher Affiliation | Collaboration | Su-Jin Shin,1 Kyungwoo Song,2 Il-Chul Moon 2 1Institute of Defense Advanced Technology Research, Agency for Defense Development 2Department of Industrial and Systems Engineering, KAIST 1,2Yuseong-gu, Daejeon, Republic of Korea sujinshin@add.re.kr, {gtshs2, icmoon}@kaist.ac.kr |
| Pseudocode | Yes | Algorithm 1 Training for Hierarchically Clustered Representation Learning |
| Open Source Code | Yes | The supplementary material is available at https://github.com/sujin6003/HCRL. |
| Open Datasets | Yes | our benchmark datasets include MNIST, CIFAR-100, RCV1 v2, and 20Newsgroups. |
| Dataset Splits | No | The paper mentions using standard benchmark datasets (MNIST, CIFAR-100, RCV1 v2, 20Newsgroups) and 'Test set performance', but does not explicitly detail the training, validation, and test splits (e.g., percentages or counts) or reference predefined splits for reproducibility within the main text. It mentions 'Supplementary Section 10 illustrates the details of the data pre-processing' but does not confirm if splits are there. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, or cloud computing instances) used for running the experiments. |
| Software Dependencies | No | The paper refers to various machine learning models and frameworks (e.g., VAE, PyTorch implicitly through deep learning models), but it does not specify any software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9, and CUDA 11.1') needed to replicate the experiment. |
| Experiment Setup | Yes | Algorithm 1 summarizes the overall algorithm for HCRL. The tree-based hierarchy T is defined as (N, P), where N and P denote a set of nodes and paths, respectively. We refer to the node at level l lying on path ζ, as N(ζ1:l) N. The defined paths, P, consist of full paths, Pfull, and inner paths, Pinner, as a union set. The GROW algorithm is executed for every specific iteration period, tgrow. After ellapsing tlock iterations since performing the GROW operation, we begin to check whether the PRUNE or MERGE operation should be performed. We prioritize the PRUNE operation first, and if the condition of performing PRUNE is not satisfied, we check for the MERGE operation next. After performing any operation, we initialize t to 0, which is for locking the changed hierarchy during minimum tlock iterations to be fitted to the training data. ... PRUNE cuts a randomly sampled minor full path, ζ , satisfying n,ζ q(ζn=ζ) < δ, where δ is the pre-defined threshold. |