At the moment Artificial Intelligence is not as intelligent as in the movies. He manages to learn a lot in a short time, and reach conclusions from what he has seen, as well as create new things, but always using information that humans have given him before.
We are still a long way from having robots “that think”, but we do have systems that are capable of understanding the relationship between objects.
This is a new machine learning system they are introducing at MIT, one that would allow a program to understand the interlocking relationships between individual objects, so that a robot will soon be able to understand things like ‘grab the spoon that’s on the left from the coffee pot and place it on top of the sink chair.
The new model represents individual relationships one at a time, and then combines them to describe the overall scene. In this way, the system manages to generate more precise images from text descriptions.
The idea is to apply it to industrial robots that have to perform complex multi-step manipulation tasks, although in the future we could have robots capable of working in more complex environments understanding the usefulness of everything around them.
The research will be presented at the Neural Information Processing Systems Conference in December. In this investigation they show how the system divides sentences into smaller parts so that they describe each individual relationship, thus modeling each part separately so that the system manages to understand the scene perfectly.
Once the scene is created, they use a machine learning technique called “energy-based modeling” to represent the relationships of individual objects.
They comment that the system also works in reverse: given an image, it can find text descriptions that match the relationships between objects in the scene.
This research has been conducted with the support of Raytheon BBN Technologies Corp., the Mitsubishi Electric Research Laboratory, the National Science Foundation, the Office of Naval Research, and the IBM Thomas J. Watson Research Center.