MEDALLION: Manifold Embedding Distillation and Attributions for Low-dimensional Latent InterpretatiON
Organizations: University of Florida, Gainsville, USA · Department of Statistics, Columbia University, New York, USA
Abstract
Dimensionality reduction is widely used to visualize high-dimensional data and guide scientific discovery, yet interpreting which original features contribute to low-dimensional representations remains challenging, particularly for black-box methods such as t-SNE and UMAP. We introduce MEDALLION (Manifold Embedding Distillation and Attributions for Low-dimensional Latent InterpretatiON), a novel framework for interpreting low-dimensional embeddings through knowledge distillation and neural feature attribution. MEDALLION distills any given embedding with a neural encoder which we explain via a broad range of existing neural feature attribution methods. Through experiments on MNIST and applications in single-cell transcriptomics, we show MEDALLION provides informative feature attributions that cast additional insights into the underlying embedding and compare its performance with existing interpretability approaches.
Figures & tables
| Unsupervised feature attribution | |||||
| Metric | Supervised reference | MEDAL + IG | KD Encoder + IG | BIR | LIMEADE |
| Pearson | MLP + IG | 0.342 (0.002) | 0.377 (0.005) | -0.010 (0.004) | -0.016 (0.006) |
| RF + MDI | 0.291 (0.004) | 0.355 (0.004) | 0.114 (0.004) | 0.04 (0.005) | |
| Spearman | MLP + IG | 0.307 (0.004) | 0.314 (0.004) | 0.028 (0.005) | 0.008 (0.007) |
| RF + MDI | 0.483 (0.003) | 0.509 (0.003) | 0.23 (0.003) | 0.113 (0.004) | |
Neural Network Interpretability
Neural network interpretability studies how models represent information and make predictions, and how to explain, test, or control their internal computations.
See all 354 papers