Find underlying ImageNet images

francis asked 5 months ago

How do I reverse find images in ImageNet which best activate on a certain node/neuron of a specif model? Eg I would like to find all images from ImageNet which activate (and have been used to train) InceptionV3/Mixed_5b/concat:44
(Would be great if you could show it either in Keras or in Lucid).
It has been done here: https://microscope.openai.com/models/inceptionv3_slim/InceptionV3_InceptionV3_Mixed_5b_Branch_0_Conv2d_0a_1x1_Relu_0/19 but it is not explained or referenced as code, how they did it.
Thanks so much in advance

1 Answers
Best Answer
Chris Staff answered 1 month ago

The technique they are using is called Activation Maximization. Here is an example using TensorFlow 2/Keras: https://www.machinecurve.com/index.php/2019/11/18/visualizing-keras-model-inputs-with-activation-maximization/

Your Answer

20 + 3 =