Really fascinated by the asymmetry between filtering negative data vs. upsampling positive data. I would have assumed the opposite but your explanation makes sense. That when combined with the alignment elasticity results makes me wonder if the effect size changes with model scale.
Really fascinated by the asymmetry between filtering negative data vs. upsampling positive data. I would have assumed the opposite but your explanation makes sense. That when combined with the alignment elasticity results makes me wonder if the effect size changes with model scale.