Evolutionary Visual Analysis of Deep Neural Networks
Key Points in Each Section
1. Introduction
A visual analytics framework for network evolution monitoring needs to possess the following three important capabilities:
- Monitor neural network evolution
- Rigorous quantitative evaluation
- Expand theoretical insights
2. Methodology
2.1 Discriminability Metric
In deep learning, the loss function is a metric for the class-wise discriminability evaluation of the neurons.
How to evaluate those neurons in the inner layers? Due to the lack of ground truth for specific classes or visual concepts, it is challenging to quantitvatively evaluate them.