CSAILVision / NetDissect

Network Dissection http://netdissect.csail.mit.edu for quantifying interpretability of deep CNNs.
MIT License
449 stars 115 forks source link

Low threshold values in the paper #28

Open Vikramank opened 2 years ago

Vikramank commented 2 years ago

Hello everyone, I would like to understand why the selected threshold value (0.04) in the paper is low? For such a low value, sometimes even random excitation can achieve that IOU right?