We now take into account the efficiency obtained by coaching this explicit illustration mannequin and examine its efficiency in opposition to that obtained by coaching supervised particular representations on real-world picture datasets. Do not forget that there was no actual information concerned whereas coaching most of the fashions.
The authors additionally in contrast the strategies on different datasets in addition to evaluated it on downstream duties as illustration studying pretraining step adopted by switch. The outcomes for this experiment are proven under
Lastly, to be able to additional perceive the properties realized from coaching on noise-like photographs, the authors additionally thought-about the visualisation of the options from AlexNet networks skilled on every of those datasets.
As might be seen from the determine above, the characteristic visualisations for these networks are as fascinating for samples skilled on these photographs obtained by noise because the options obtained for real-world photographs
All these evaluations counsel that even when one doesn’t have entry to real-world information, one can nonetheless be taught a significant illustration simply by acquiring fashions by way of structured procedures that pattern noise.
As is clear by way of out the article, all of the work mentioned is introduced within the following fascinating work and a few of its references.
The code for this work is shared by the authors on the following github hyperlink: https://github.com/mbaradad/learning_with_noise
