Motor Cortex Neural Recordings
Neuroscientists recorded spike trains from 200 neurons in motor cortex while monkeys performed reaching movements to 8 different target locations. Each trial is labeled with the target direction. The team has 1000 trials of data and wants to build a decoder to predict movement direction from neural activity.
Research Question:
Which technique would be most appropriate for finding a low-dimensional representation to use in the decoder?
Tip: Don't worry about getting things “wrong”—this exercise is about learning what outcomes different techniques produce in different situations. Try selecting various techniques to understand their practical implications!
Select a Technique:
Principal Component Analysis
PCA finds directions of maximum variance without considering class labels. Can create phantom oscillations with nonlinear data.
Click to learn more and see if this is a good choice
Linear Discriminant Analysis
LDA finds dimensions that maximize separation between known classes while minimizing within-class variance.
Click to learn more and see if this is a good choice
Independent Component Analysis
ICA separates mixed signals into statistically independent components.
Click to learn more and see if this is a good choice
Uniform Manifold Approximation and Projection
UMAP primarily preserves local neighborhoods in high-dimensional data; distances between clusters may not be meaningful.
Click to learn more and see if this is a good choice
Show Hint