I will share some of our recent works where machine learning (ML) was used to obtain a novel view of galaxies. Firstly, we report the discovery of a two-dimensional Galaxy Manifold within the multi-dimensional luminosity space of 11 bands that span from far ultraviolet to near-infrared for z<0.1 galaxies. The found manifold suggests a paradigm where galaxies are characterized by their continuous two-dimensional space as opposed to the traditional view of two galaxy groups. Secondly, we propose a data-driven model of galaxy star formation histories (SFHs). Extracting a galaxy’s SFH from its observed spectral energy distribution (SED) is challenging. Therefore, SFHs are commonly approximated with parametric forms that impose strong priors on their shape and property measurements. Physically-motivated SFH models from simulations may then be favored. Instead of building an extensive library of simulated SFHs, we propose using a recent ML development called generative modeling. These models learn the intrinsic distribution of SFHs from simulations and can generate new physically-motivated SFHs, which can be matched to observations. The advantages of such methods are bidirectional, to constrain the galaxy property measurements from observed SEDs, and to constrain galaxy formation models from observations.