Dynamics of the accelerated t-SNE
Authors: Kyoichi Iwasaki, Hideitsu Hino
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through rigorous theoretical analysis and empirical validation, we show that our approach offers a further insights into the dynamical properties of t-SNE. Our main contributions are summarized as follows: We explore an approach of dynamical system analysis of t-SNE including a momentum term, and further extended to the NAG. ... We numerically evaluate our results by comparing the t-SNE algorithms with and without acceleration and their ODE counterparts with real-world datasets: KDDcup1999 and MNIST. |
| Researcher Affiliation | Academia | Kyoichi Iwasaki EMAIL The Graduate University for Advanced Studies, SOKENDAI; Hideitsu Hino EMAIL The Institute of Statistical Mathematics |
| Pseudocode | No | The paper describes mathematical equations and theoretical derivations for the dynamics of t-SNE, along with descriptions of algorithms (GD, MM, NAG), but it does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | Although the term acceleration often suggests computational speed-up, we clarify that our current framework focuses on theoretical acceleration in the continuous-time dynamics, rather than practical runtime efficiency. In fact, due to the reliance on full eigendecomposition of the graph Laplacian, the overall computational complexity remains O(n3) in our current implementation. There is no explicit statement or link indicating that the code for the described methodology is publicly available. |
| Open Datasets | Yes | We numerically evaluate our results by comparing the t-SNE algorithms with and without acceleration and their ODE counterparts with real-world datasets: KDDcup1999 and MNIST. ... KDDCup1999: https://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html, CC BY 4.0 license referenced by https://archive.ics.uci.edu/dataset/130/kdd+cup+1999+data; MNIST: https://scikit-learn.org/stable/modules/generated/sklearn.datasets.fetch_ openml.html, BSD license provided by sklearn; Olivetti Faces: https://scikit-learn.org/stable/modules/generated/sklearn.datasets.fetch_olivetti_faces.html, BSD license provided by sklearn |
| Dataset Splits | No | For the KDDCup1999 dataset... We extracted 100 samples of data for each of the 5 labels ( smurf , neptune , normal , back , satan ) from the dataset, totaling 500 samples. ... for MNIST dataset... we focus on 400 images with 4 labels( 2 , 4 , 6 , 8 ) totaling 1600 samples... The paper describes data selection for visualization but does not specify train/test/validation dataset splits or cross-validation methods. |
| Hardware Specification | Yes | The experiments were conducted on a laptop PC with a 12th Gen Intel(R) Core(TM) i7-1255U processor, 500GB storage, 16GB memory, using Python. |
| Software Dependencies | No | The experiments were conducted on a laptop PC with a 12th Gen Intel(R) Core(TM) i7-1255U processor, 500GB storage, 16GB memory, using Python. The paper mentions Python as the programming language but does not specify its version or any specific library versions used. |
| Experiment Setup | Yes | In this paper, we set perplexity = 30, as in Cai & Ma (2022). In all experiments in this section, all initial embedding vectors were generated randomly. ... The optimization is done with step size parameter h = 5, momentum parameter m = 0.5, Perplexity is 30, and exaggeration parameter α = 10. ... We propose a stopping criterion based on a threshold (e.g., 0.01) and demonstrate its effectiveness in subsection 8.2. |