This saturday Marco gave an invited talk at the THE 29TH INTERNATIONAL CONFERENCE ON COMPUTATIONAL LINGUISTICS.

Summary :

Deep nets are great, and they would be even greater if they could collaborate with each other through a flexible communication protocol. Since many networks come equipped with language model interfaces--that is, they can process and produce tokens in English (or other languages)--it makes sense to look into natural language as a universal neural network communication protocol. The questions I would like to discuss in my talk, then, are the following:

• What are the challenges we face in developing a universal language interface among deep networks?
• What do we currently know about the way in which deep nets are using language to interact in machine-to-machine communication scenarios?
• Which characteristics of human linguistic communication do we expect or wish machine-to-machine communication to possess? Should it be essentially the same as human-to-human (and human-to-machine) communication?