Learning Latent Temporal Manifolds For Recognition And Prediction Of Multiple Actions In Streaming Videos Using Deep Networks