Autoencoder Interactive Demo
The demo may take a second or two to load. Scroll down for more info on how the demo works. For a more detailed explanation, check out this page.
What am I looking at?
What are those sliders?
This demo is split into two tabs: image reconstruction and image interpolation.
In the image reconstruction tab, an image from the dataset is displayed on the left and the autoencoder’s reconstruction of the image is displayed on the right. Hitting “New Image” will randomly select new images for the autoencoder to reconstruct.
In the image interpolation section, two images are chosen from the dataset and put on the left and right. The big slider mixes between the left and right image and displays the result in the center. You’ll also see the sliders at the bottom update with the big slider.
The sliders at the bottom each represent some trait of the image that the autoencoder thought was important. You can think of them as key descriptors, like how you might describe an apple as “red, round, and shiny.” Each slider corresponds to a unique descriptor used by the decoder.
The sliders are sorted in order of importance, so the first slider causes larger changes than the last one. You can use the sliders to directly change the values of these key descriptors, but you will notice that it’s pretty hard to know exactly what each slider actually “means” to the model.