I'm building upon insights from this paper (https://arxiv.org/pdf/2403.03950.pdf https://arxiv.org/pdf/2403.03950.pdf) and believe that classification can somet

I'm building upon insights from this paper (https://arxiv.org/pdf/2403.03950.pdf https://arxiv.org/pdf/2403.03950.pdf) and believe that classification can sometimes outperform regression, even when dealing with continuous output values. This is particularly true in scenarios where the output is noisy and may assume various values (multi modal). By treating the problem as classification over discrete bins, we can obtain an approximate distribution over these bins, rather than settling for a single, averaged value as regression would yield. This approach not only facilitates sampling but may also lead to more favorable loss landscapes. The linked paper in this comment provides more details of this idea.