The impact of trial-to-trial variability in activity on estimates of manifold geometry.
<p>We simulated structured concept manifolds formed by sets of activity patterns with varying amounts of trial-to-trial variability, which we refer to here as “measurement noise”. We estimate geometric properties of the manifolds as a function of the measurement SNR (, x-axis), which is define...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>We simulated structured concept manifolds formed by sets of activity patterns with varying amounts of trial-to-trial variability, which we refer to here as “measurement noise”. We estimate geometric properties of the manifolds as a function of the measurement SNR (, x-axis), which is defined as the variance of the average activity across images, divided by the average variance of measurement noise. We consider both correlated (solid curves) and uncorrelated (dashed curves) measurement noise. We also show values of geometric properties inferred from trial-averaged activity with 1, 3 and 5 samples (colors) for each exemplar. The ‘true‘ (asymptotic) values of the geometric properties (solid magenta lines) are achieved when measurement noise tends to zero, i.e., when measurement SNR is large. Estimates also tend toward their asymptotic values as the number of samples available for trial-averaging increases. As shown in <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1013416#pcbi.1013416.s004" target="_blank">S4 Fig</a>, we find that all estimates of geometric properties based on encoding model outputs are consistent with the ‘denoising‘ effect of an increased number of samples (this effect is illustrated schematically by the dashed magenta line in the leftmost panel). In other words, estimates of geometric properties based on encoding models yield higher estimated values for all properties except Dimensionality, which is exactly the effect of increasing the number of samples for trial averaging. Thus, the encoding model we use here impacts our estimates of geometric properties in the same direction as averaging across increased numbers of samples.</p> <p>(TIF)</p> |
|---|