Question
Laurens van der Maaten developed a MATLAB “toolbox” that collects over 30 techniques for this action, including Maximum Variance Unfolding and Local Tangent Space Alignment. The “swiss roll dataset” is a toy example on which this action can be performed to unravel it. Somewhat surprisingly, the Barnes-Hut method for galaxy simulation can substantially speed up a technique for this action called t-SNE (“T-snee”). It’s not compression, but autoencoder neural networks can perform this action due to having very small hidden layers. Nonlinear techniques for doing this action are also known as (*) manifold learning. Principal component analysis performs this general action on data to avoid a certain “curse” and make it easier to visualize on 2 or 3-D plots. For 10 points, name this action that turns an input data point into an output with a smaller number of coordinates. ■END■
ANSWER: dimensionality reduction [accept descriptions of mapping or embedding high-dimensional data to fewer dimensions; accept manifold learning until read; accept visualizing high-dimensional data; prompt on “clustering” with “what action that can result in clustering?”; prompt on “embedding” before read with “embedding data with what property?”; prompt on “visualizing data” with “what property does the data have?”; prompt on “projection” with “what desired effect does projection have?”]
<BC>
= Average correct buzz position
| Conv. % | Power % | Average Buzz |
|---|
| 100% | 75% | 75.25 |
Back to tossups