Suggested improvements in DMM based AI

[SolarpunkFan]

AI Advancement In Answering Questions About Text, Images

Slowly but surely, artificial intelligence (AI) is advancing. Where it stands in comparison to human intelligence is difficult to say because people excel across a broad set of intellectual tasks, while AI tends to be narrowly focused. But machines keep getting better at tasks like sifting through vast quantities of data, understanding natural language, and recognizing objects in images.

Over the weekend, scientists from a Silicon Valley company called MetaMind described their efforts to advance the state of the art in a research paper, "Dynamic Memory Networks for Visual and Textual Question Answering."

The paper explains improvements in memory and input modules for a system called a dynamic memory network (DMN), a general architecture for answering questions about text and images. A DMN might, for example, parse text describing a series of events in which a person passes through a series of rooms and drops an object in one of them. If asked about the location of the object, the DMN could infer the correct answer from the story.

MetaMind's researchers have devised a way to have their data model learn the facts it needs to reason without labeling those facts in a training session. Their approach involves creating a component that considers text in a bidirectional manner, rather than analysis that flows one way from beginning to end.

By allowing both past and future sentences to be considered when attempting to reason about a specific situation, the researchers see more accurate results. The technique also can be used to improve the identification of objects in pictures.

The whole article is at the site linked. Thoughts about this? It's really amazing how fast the advances in AI are coming along.

[Starglider]

Thoughts about this? It's really amazing how fast the advances in AI are coming along.

Generally yes but this is not really an example of that. Other approaches have been able to do this kind of reasoning since the late 1970s. It is news only in the sense that connectionist approaches are slowly managing to replicate capabilities of symbolic approaches, thirty years later. The idea is of course that connectionist approaches won't hit the scaling problems that the symbolic ones did; speculative, because connectionist approaches constantly hit brick walls as well. Recently connectionist approaches have had more breakthroughs that overcome walls that have been around for a while, but I think this is partly due to the current distribution of research dollars; connectionist approaches got / are getting a lot of funding derriving from the big data analytics boom.

[SolarpunkFan]

Generally yes but this is not really an example of that. Other approaches have been able to do this kind of reasoning since the late 1970s. It is news only in the sense that connectionist approaches are slowly managing to replicate capabilities of symbolic approaches, thirty years later. The idea is of course that connectionist approaches won't hit the scaling problems that the symbolic ones did; speculative, because connectionist approaches constantly hit brick walls as well. Recently connectionist approaches have had more breakthroughs that overcome walls that have been around for a while, but I think this is partly due to the current distribution of research dollars; connectionist approaches got / are getting a lot of funding derriving from the big data analytics boom.

Ah, that makes sense. Sadly I can't say I'm versed in the different approaches in A.I. but I do understand the gist of your post.

[Zixinus]

Starglider has written a rather extensive thread about this, discussing various approaches: http://bbs.stardestroyer.net/viewtopic. ... 9&t=136633