Perplexity-free Parametric t-SNE
This work addresses a specific problem in dimensionality reduction for researchers and practitioners using t-SNE, offering an incremental improvement by extending parametric t-SNE to be perplexity-free.
The paper tackles the limitation of parametric t-SNE being bounded to a user-defined perplexity parameter, which restricts dimensionality reduction quality, by proposing a multi-scale parametric t-SNE scheme that eliminates perplexity tuning and uses a deep neural network for mapping, resulting in reliable embeddings with out-of-sample extensions that are competitive with the best perplexity adjustments in terms of neighborhood preservation on multiple datasets.
The t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm is a ubiquitously employed dimensionality reduction (DR) method. Its non-parametric nature and impressive efficacy motivated its parametric extension. It is however bounded to a user-defined perplexity parameter, restricting its DR quality compared to recently developed multi-scale perplexity-free approaches. This paper hence proposes a multi-scale parametric t-SNE scheme, relieved from the perplexity tuning and with a deep neural network implementing the mapping. It produces reliable embeddings with out-of-sample extensions, competitive with the best perplexity adjustments in terms of neighborhood preservation on multiple data sets.