Sliding down the stairs: how correlated latent variables accelerate learning with neural networks