filmov
tv
Neural Program Synthesis

Показать описание
Денис Ракитин, НИУ ВШЭ
The problem of program learning consists of generating a computer program consistent with some specification. Contemporary studies showed that using neural program synthesis approach for solving this type of task can result in a more robust and less data dependent algorithm comparing to the classical methods. Moreover, adding a program synthesis module as a building block of another neural-based algorithm can provide it with prior structural knowledge and make its performance more interpretable. In this talk I will make an introduction to this approach provided with some examples of usage. Main part of the speech will be in covering 2 recent papers that successfully combine neural program synthesis with deep representation learning in application to visual question answering problem.
The problem of program learning consists of generating a computer program consistent with some specification. Contemporary studies showed that using neural program synthesis approach for solving this type of task can result in a more robust and less data dependent algorithm comparing to the classical methods. Moreover, adding a program synthesis module as a building block of another neural-based algorithm can provide it with prior structural knowledge and make its performance more interpretable. In this talk I will make an introduction to this approach provided with some examples of usage. Main part of the speech will be in covering 2 recent papers that successfully combine neural program synthesis with deep representation learning in application to visual question answering problem.