Choose a photo
Pick a JPEG or PNG of a skin lesion, or start from one of the example images, and run the model on it.
K-nearest Index of Typicality in Hidden layers
Machine metacognition · Research demo
Checking whether a model has seen images like yours
KITH is a research demo built around a computer model that sorts close-up photos of skin lesions into benign (harmless) or malignant (cancerous). It learned from 5,772 labelled photos.
Give it a photo and you get two results: the model’s answer with its output probability (the probability the network gives its top answer), and a familiarity score showing how closely your photo resembles the photos it learned from.
Your photo stays on your computer. Nothing is uploaded.
Explore eight worked examples →Read how it works →
A research demo. Not medical advice: if a lesion on your skin worries you, see a doctor.
Measuring Neural Network certainty by distance to known examples
Prediction
Which label does the model choose, and how sure is it?
Familiarity
How similar is this image to the training photos?
Illustration only, not real results.
Combating over-confidence in neural networks.
Read two results
See which label the model chooses, and how similar your image is to the training images it most resembles.
Check familiarity at different stages of the network
Try an image to compare the familiarity score at the early stages of the model with the score at the late stage.
The research question
Models like this one give an answer without saying how they reached it. Adding a form of metacognition, a sense of how well the model knows what it is looking at, could help a model improve its own answers and help it tell people how it arrived at them. Our familiarity score is one attempt at this. It measures how familiar a photo looks to the model compared with the photos it learned from. An unfamiliar photo is not always a problem, but it can be critical when a model is sure but wrong.