Recurrent Neural Networks for Video Similarity Metrics and Video Captioning.

The objective of this project is to develop a system that provides video recommendations to the user based on the content of the video they are currently watching. Existing methods generally create their recommendations based on the title or tags of the video or on the viewing patterns of other users who have watched the same video. However, text based methods are only accurate if the titles and text are correctly spelled, use similar colloquial language, and accurately reflect the important content of the video, which are factors that are not guaranteed. Furthermore, it is often difficult to infer meaningful viewing patterns and determine similar user groups when attempting to create video recommendations based on the viewing behavior patterns of similar users, particularly if the user base is small. Hence, a more accurate system for video recommendations could be created through an artificial intelligence method that is able to autonomously infer what the important content of a video is. The method would then use this important content to find other videos that contain similar important content in order to create video recommendations for the user as well as describing this important content in a natural language sentence.

Faculty Supervisor:

Richard Zemel

Student:

Rohan Chandra

Partner:

Vemba

Discipline:

Computer science

Sector:

Information and communications technologies

University:

University of Toronto