We develop a transferable machine learning model which predicts structural relaxation from amorphous supercooled liquid structures. The trained networks are able to predict dynamic heterogeneity across a broad range of temperatures and time scales with excellent accuracy and transferability. We use the network transferability to predict dynamic heterogeneity down to the experimental glass transition temperature, $T_g$, where structural relaxation cannot be analyzed using molecular dynamics simulations. The results indicate that the strength, the geometry and the characteristic length scale of the dynamic heterogeneity evolve much more slowly near $T_g$ compared to their evolution at higher temperatures. Our results show that machine learning techniques can provide physical insights on the nature of the glass transition that cannot be gained using conventional simulation techniques.