learn/model families
Dimensionality reduction
Dimensionality reduction is a way of describing data with fewer numbers than it arrived with, choosing new coordinates that keep the patterns scientists care about and discard the rest.
What it is
Scientific measurements often come as long lists of numbers. One material might be described by a few hundred numbers, one protein by thousands. Each number is a dimension, and data with many dimensions is hard to plot, hard to compare and hard to learn from. Dimensionality reduction replaces those many numbers with a handful of new ones. The new numbers are not the original measurements but combinations of them, worked out so that items which were close together in the long description stay close together in the short one, and items which were far apart stay far apart.
Methods differ in what they try to preserve. Some keep the broad spread of the whole dataset. Others, often called manifold methods, assume the data lies on a curved surface inside the larger space and try to keep each point's local neighbourhood intact. Some are built mainly for drawing a two-dimensional picture to look at. Others produce a compact description meant to be fed into a later model. The number of new dimensions is a choice, and it is usually settled by testing how well the shortened version supports whatever comes next.
Why scientists use it
Fewer dimensions make work possible that otherwise is not. Models trained on very long descriptions need far more examples before they can tell signal from coincidence, because there are so many ways to fit the data by accident. Shortening the description eases that. It also cuts computing time, and it removes the repetition that comes when several measurements say nearly the same thing.
The other reason is simply to see. Humans cannot picture a hundred dimensions, but they can look at a scatter plot. Reducing data to two or three dimensions lets researchers notice that samples fall into families, that one batch sits apart from the others, or that a property varies smoothly across the picture. This is why reduction and clustering are often used together: reduction changes the coordinates, clustering then decides which points belong in the same group.
In the record
Limits
Something is always thrown away, and there is no guarantee it was the unimportant part: a difference that matters for one question can vanish in coordinates chosen for another. Pictures from neighbourhood-preserving methods are especially easy to over-read, because the size of the gaps between groups, and sometimes the groups themselves, can be artefacts of the settings rather than features of the data. The new coordinates usually have no physical meaning, so they cannot be interpreted as a measured quantity, and results can shift when the method is rerun or retuned.
Written by AIxSci: the background from general knowledge, the examples from checked records.
By field
- Dimensionality reductions in structural biology and protein science
- Dimensionality reductions in materials and computational chemistry
- Dimensionality reductions in astronomy and astrophysics
Every article
aix-00276
Machine learning sorts a simulated polymer into ordered and disordered beads
aix-00204
Neural networks stand in for a slow model of interstellar gas clouds
aix-00181
Protein language model embeddings used to predict succinylation sites in proteins
aix-00186
Language model trained on protein sequences predicts shape and behaviour of disordered proteins
aix-00125
Classifying stars by type from a single wide-band telescope image
aix-00100
Astronomers test pre-trained image models for sifting telescope alerts
aix-00088
Neural network retrained to predict how crystals absorb light, from few examples
aix-00082
Neural networks turn simulated dark matter maps into mock X-ray and microwave cluster images
aix-00081
Language models trained on protein sequences used to predict sugar-attachment sites
aix-00066
Machine learning sifts 25,000 known materials for quantum-technology host candidates
aix-00060
Neural networks stand in for slow cosmology code in parameter fitting
aix-00049
Eight machine learning methods sorted supernova gravitational wave signals by nuclear physics model
aix-00042
Two unsupervised methods compress fast radio burst spectra into a handful of numbers
aix-00040
Statistical model of 5,441 perovskite solar cells predicts efficiency and suggests recipes
aix-00001