Ha, I remember the poster from the conference, it was quite crowded when I passed by. This one seemed to have a big focus on data augmentation in the DCT space.

Ha, I remember the poster from the conference, it was quite crowded when I passed by. This one seemed to have a big focus on data augmentation in the DCT space. I was asking myself (and the author) whether you couldn’t eke out a little more efficiency by trying to quantize your network similarly to the default JPEG quantization table. As I understood, currently all weights are quantized uniformly, which does not make sense when your inputs are heavily quantized, does it? Maybe I should dive a little deeper into the Uber paper, they were focusing a bit more in the quantization part. Sorry if I’m talking nonsense, this is absolutely not my area, but I found the topic captivating.