We predict the occurrence of nucleation in the two-dimensional Ising model using the Convolutional Neural Network (CNN) and two logistic regression models. CNN outperforms the latter in systems with different interaction ranges and sizes, especially when the size of the system becomes large. We find that the CNN decreases its prediction power as system gets closer to the spinodal. We give explanation using the ramified droplet structure predicted by spinodal nucleation theory.