eb1e78328c46506b46a4ac4a1e378b91-AuthorFeedback.pdf

Itistherefore7 striking that for allthe other models, the representations cluster soclearly into orange and blue inthe plots from the8 paper -suggesting significant differences between representations learned from pretraining vsrandom initialization.9 Referenced paper: Thank you for the reference to the paper on developing CNNs for tuberculosis screening and12 visualization, which we will add tothe related work. Inparticular,thepaper14 states that'the use of pretraining is outside the scope of the paper' - and does not compare the effect of pretrained15 weights vs random initialization. This is the core of the transfer learning question, and a central part of our paper.16 This phenomenon is consistent with related effects we have found since the submission in papers that36 we reference inthe updated version of the paper: [6] Neural Tangent Kernel, (Jacot etal), [7] Are all layers created37 equal?(Zhangetal).