3 ms·
Dialogue as Dataflow: A new approach to conversational AI
- ilaksh 6y agoMy rough understanding is that a key part of this is using neural networks to build and update graphs or programs representing tasks. The level of emphasis on composition might be pretty new. But I think that building and updating highly compositional graphs or programs driven by neural networks interpreting input is an incredibly powerful general idea that may be applied to a wide variety of AI challenges. One thing I have been thinking about in computer vision is trying to create a scene graph or maybe scene description program by using neural networks to interpret visual input. Then running that program and comparing the output scene to the input to see how accurate the scene description program is. Seems incredibly hard, but I want to believe something along those lines could work if there is first enough curriculum training of subprograms that could be recognized from scene elements in the input.