Disentangled Features

Features are disentangled when Neural Network learns features in a way such that each Unit detects a specific real world concept.

For example in CNNs when channel 394 might detect skyscrapers, channel 121 dog snouts, channel 12 stripes at 30 degree angle.

This property would make it extremly easy to interpret Neural Networks by just looking at the most activating Unit and mapping it to its concept.