In this thesis we apply likelihood maximization methods as well as mean-field approximation to construct equilibrium and non-equilibrium Ising models for neural activity in the form of calcium imaging recordings. Before applying these methods to the neural data, we carried out tests on samples generated from the models themselves. We also investigated the statistical properties of the neural recordings themselves, as well as the performance of these inference methods under various circumstances.Our findings indicate that both the equilibrium and non-equilibrium Ising models are capable of capturing the essential statistical features of the calcium imaging recordings, not only reproducing constraint observables but also predicting statistical properties of the neural activity not used to fit the models. Additionally, our findings show that, for these particular calcium recordings, the mean-field methods perform better with wider bins. And while the mean-field methods appear to break down for large numbers of neurons, the exact likelihood maximization procedures for these two models are reliable also for larger system sizes.