Adaptive deep learning for efficient media content creation and manipulation

Abstract

Automatic editing of media compositions is performed using media editing applications equipped with neural-network-based deep-learning models. The automatic editing is adapted to the practices of local users of a media editing application by training the models on a combination of media compositions previously edited by third-party media editors and media compositions edited by local users. Training data input vectors for the model comprise representative portions of a composition's raw media, and corresponding output vectors include values of parameters that define editing functions applied to the raw media to generate an edited media composition. A user interface enabling a user to adjust and monitor machine learning parameters is provided. Adaptive automatic editing may assist in the creation of video and audio compositions, as well as in the generation of musical scores.