Performance¶
WIP
This is a work in progress. refresh the page to see any recent changes made
t-SNE visualization of the exemplars¶
t-SNE is a dimensionality reduction technique that is particularly well suited for the visualization of high-dimensional datasets. It is a nonlinear technique that is particularly well-suited for embedding high-dimensional data into a space of two or three dimensions, which can then be visualized in a scatter plot. Specifically, it models each high-dimensional object by a two- or three-dimensional point in such a way that similar objects are modeled by nearby points and dissimilar objects are modeled by distant points with high probability.
Here is an example of a t-SNE visualization of the exemplars in the UAV project dataset. Further apart exemplars are more dissimilar, while closer exemplars are more similar. Separation is a good indicator of the quality of the exemplars, so we can see here that the exemplars are very well separated.
Multi-class precision-recall curves¶
The Multi-class precision-recall curves show the precision/recall as you move through different decision thresholds.
Curves that hug the top-right corner are generally good - any decision threshold has good precision/recall.
The AP number is the area under the curve for a particular class which is reported in the table in
this example, is a useful metric to see how well the model performs on a particular class.
In this example, Mesodinium (AP=0.62) has moderate performance and Akashiwo (AP=1.00) is perfect.
Click on the image to see a larger version.
Confusion matrix¶
The confusion matrix is useful to see where there is confusion between classes. Many times these confusions are due to labeling errors, but sometimes this is also due to simply closely related classes that are difficult to distinguish.
Click on the image to see a larger version.
🗓️ Updated: 2025-10-11