The quantization used for JPEG is optimized to throw away information in the frequency space that doesn't matter much to human perception, but I wonder if that

The quantization used for JPEG is optimized to throw away information in the frequency space that doesn't matter much to human perception, but I wonder if that is also optimal for training neutral networks? Also, as far as I know, the human eye doesn't process images in blocks. I also wonder how blockless encoders such as JPEG 2000 would fare in this approach.