To develop a multimodal DNN, while also incorporating input about movie poster features learned in a data-driven fashion, for movie box office revenues prediction before the theatrical release of the movie. The author argues that it is expected that the extracted features from movie posters may have an impact on the performance of the DNN.

1- A CNN would be trained first on learning and extracting features from movie posters.
2-Then, CNN is used as a module in the multimodal DNN. The movie poster content extracted by the CNN is used with other movie-related data as input.

Using this model, audio and video data from movies can be used as inputs by building additional multimodal DNNs.

Accuracy results show a good sign of the use of multimodal DNN as a prediction system, however, accuracy can be increased by adding more layers to the CNN to help extract more features from the movie posters but up to a certain extent to avoid over fitting.