Neural networks will guess what you will say in a conversation or text.

We have seen robots interact with humans in both science fiction and real world stories. But artificial intelligence is moving to another level: it will be able to guess what you are about to say. And this is already true thanks to a work done by scientists from the University of Michigan, in the United States.

They developed a mechanism based on the Reservoir Computing (RC) paradigm, which is based on a neural network, capable of processing varied data. The latter was conceived from memristors, electronic devices that “record” in their memory something or some pattern that has happened before.

The construction of a neural network can be understood as the neurons of a brain, consisting of connections between nodes. In order to be able to make this kind of prediction, she is usually submitted to a large number of questions accompanied by their answers. With the completion of this training, you also become able to recognize something you do not yet know.

However, with the use of memristors in the computer system, a considerable part of this process is excluded, since its reservoir does not require training. Thus, to make predictions, the reservoir itself takes care of selecting and simplifying the most important data.

They are then passed on to a second network, which only requires training with the simplest neural networks. Therefore, it will process them by selecting the ones that are considered most appropriate, until reaching the best possible level, that is, with the least possibility of errors.

Initially, the neural network can only predict words and numbers that will still be written in a text. In this case, the experts were able to “train” the neural network with 88 memristors, reaching a high accuracy of 91.1%. This number of components was as close to ideal as increasing this number to 112 units would increase the rate to just 91.5%.

Given this, experts say that the system can still be optimized with some adjustments during the preprocessing of this information. That done, they can move on to the next step, where they want to make you able to recognize words early in a spoken conversation between two people.

This technology should benefit several areas, especially those that need speech analysis and action recognition / prediction tools. In addition, they can be used to increase the efficiency of artificial intelligence in general and machine learning and neuromorphic computers, which work similarly to the human brain.

Neural networks will guess what you will say in a conversation or text via TecMundo