Insightful Saliency Maps

Explore the saliency maps of stylized images. See how they compare to the saliency map of the original, un-stylized image.

A saliency map identifies what regions in an image are the most important to a neural network's prediciton. XRAI is a saliency method that uses regions and colors to interpret how a computer understands an image. Typically, the image is split into a gradient of colors that represent what features in the image are most important to identifying the class of the image all the way to what features are least important to identifying the class of the image.

1 Choose a stylized image:

Dog x Tiger Dog x Zebra Dog x Trucks Dog x Bikes Dog x Elephant Bear x Trucks Bear x Tiger Bear x Zebra Bear x Bikes Bear x Elephant

Prediction: ; % confidence

40%

Show XRAI Hide XRAI Overlap XRAI

Original image:

Dog Bear

Prediction: ; % confidence

40%

Show XRAI Hide XRAI Overlap XRAI

Visualize the similarity between the two saliency maps.

Intersection:

15 delta

Comparison of Colors:

6 colors

Compute the average saliency map

1 Choose an Image Class:

Dog Bear

2 Select the images below to include in the average:

Generate Average Saliency Map

Average Saliency Map:

Original Saliency Map:

Draw what the AI sees.
See how your perception compares to the AI.

The AI makes predictions based on regions it feels are important to identifying the image. Can you guess what places in this image the AI feels are the most important to identifying it?
What is important and somewhat important? Finally, what is least important?

Draw on the image how you think a computer would classify it

Your browser does not support the HTML5 canvas tag.

Select a color

Yellow Orange Red Purple

Percentage of pixels correct so far

= Here's a hint! This is how the computer sees

Here's what the computer sees

Your browser does not support the HTML5 canvas tag.

I'm done! Show me how an AI sees this image

Compare accuracy of different models through saliency maps. Is there one that consistently performs better than others?

Model Comparison:

Inception V3 Intersection:

15 delta

VGG 16 Intersection:

15 delta

ConvMixer Intersection:

15 delta