[Lesson 1] Bird/Forest classifier predicts "Bird" for unrelated images

Part 1 2022
1

I’m seeing some unexpected behavior with my binary classifier trained to distinguish between birds and forests.

When I feed it images that contain neither (e.g., a photo of a car or a book), the neural network consistently classifies them as a bird with over 99% confidence.

Could someone help explain why the model is failing so confidently on out-of-distribution inputs, and how I can prevent or handle this behavior?

1 Like
2

You can take a look at fasttransform: Reversible Pipelines Made Simple – fast.ai

It explains the usage of Interpretation class which helps understand what your model actually “sees” before classifying something.