5 ms·
I looked through the GAN paper that this idea uses, and it seems that they generate training data by downsampling high res images. So you can presumably create
by anjc 7y ago
I looked through the GAN paper that this idea uses, and it seems that they generate training data by downsampling high res images. So you can presumably create large amounts of high quality training data for problems like this.
Do such methods introduce any problems in terms of e.g. overfitting (say, overfitting to the method that you use to degrade images)? Also, do fields tackling ML problems such as this where you can generate perfect training data tend to have faster improvements over problems where you can't (i.e. where you need to manually label training data, or where the target is ambiguous like sentiment analysis)?
- ivymike3257 7y agoHey thanks for the question. Regarding your first question: that was exactly what I wondered, and I found out that they don't necessarily overfit to the type of degradation (the repo shows a section called "extreme super resolution" where the input is already high res, and the output shows no artifacts). But they do tend to overfit to the domain of images. So you would probably observe artifacts running this on pictures of...curcuit boards? But those are also functions of batch norm, as some literature points out, and I tried to mitigate this problem by not using batch norm layers in the end of the network. I always welcome people testing it out and reporting any problems they observe though. For the second question: I would say definitely. You don't need to waste a lot of time worrying about data quality, or even annotation in this case. Also, ground truth to the model is actual, real, without noise ground truth, so they may be able to learn better features representative of the distribution of data. I'm sure there's an argument to be made here about the generalization capability of models being trained on noisy data being higher, but I personally think cleaner data = better data. Especially for a GAN where the goal is to learn the distribution of your input space. Noise would only hinder that.
- mark_l_watson 7y agoYou are right that this class of problems has the advantage of cheaply getting a lot of training data. Another type of model with the advantage of cheap training data is BERT like models for natural language processing where input training data is just natural language text with random words replaced by ‘holes’ and target output is the original text with no missing words.
- alexfromapex 7y agoYeah there is definitely data leakage using that method.