DeepMind’s Artificial Intelligence can apply learned information to complete new tasks

Google’s roofing company Alphabet’s artificial intelligence platform DeepMindcan learn to generalize beyond its sudden experience. In a study conducted jointly with Stanford and University College London, DeepMind scientists investigated how systems can apply the information they learn on one task to other tasks. The teams conducting the study reported that DeepMind’s artificial intelligence correctly used the nature of a language to interpret instructions that had never been seen before.

Scientists say that although artificial intelligence systems trained in ideal or reduced situations fail to demonstrate a compositional or systematic understanding of their experiences, this competency is a wide range of multi-faceted, multi-faceted it can easily occur when they can access many examples with observations. The team believes that the ability to entertain an idea implies the ability to entertain thoughts with semantically related content; he investigated the extent to which they could give a systematic model of artificial intelligence, a cognitive concept. For example, if you’re in When it comes to systematics, it’s possible that a person who understands“John loves Mary”mightthink, “Mary loves John.”

Given two objects that are randomly positioned and trained using rewards to strengthen desired behavior, agents learn the concept of“removal, upgrade”to apply to objects they have never seen before. they reported it. In addition, according to the instructions, the next task, which requires them to place objects on beds or trays, is said to have achieved 90 percent placement accuracy.

In a separate test, the team investigated tasks that could be solved without relying on language or language. They positioned the agent on a virtual grid of eight randomly positioned objects. Object types have been correctly identified, and agents have received a reward for collecting such objects.

As a result, researchers have proven that three factors are critical in all tests. The number of words and objects experienced during training, the first-person perspective and the variety of inputs provided by the agent’s perspective over time.

Categories:   Technology