Researchers from The University of Michigan came up with a way to make a new generation neural network. AT her basis will lie memristory, so that the neural network can anticipate the actions or words of a person.
Today's neural networks are designed with application of optical technologies. This creates a bunch of restrictions, such as the large amounts of time it takes to study. For example, today's neural network wrote a steep tale of Harry Potter, her at For several weeks or months, you will need to feed pages from original novel, and then a long time to learn to understand the context. The time spent is proportional to the complexity of the task.
Memristors It is a special type of material that can and perform logical operations, and store data. This will allow you to drop most of the training process so the tank with data, which is the most important component of the system, requires training. Due to the combination of the two functions, it can recognize the context of the data received and can immediately transfer them at simplified form for more specific training of second-tier networks.
Not know why, but developers of the system plan to use it for predicting human speech. The There is, after a certain training, to predict the words of a person for second to said. Probably, to quickly display it from yourself.
Read in Russian: Ученные придумали, как научить нейросеть предсказывать будущее