Learning binary latent variable models: A tensor eigenpair approach