Hybrid computing using a neural network with dynamic external memory
Nature 2016
Original link:http://www.nature.com/nature/journal/vaop/ncurrent/pdf/nature20101.pdf
absrtact : AI Neural Networks have been very successful in perceptual processing, sequence learning, reinforcement learning, but limited to their ability to represent variables and data structures, and the ability to store knowledge for long periods of time because of the lack of an additional memory unit. Here, we introduce a machine learning model called: A differentiable neural computer (DNC), which contains a neural network that can read and write an additional memory matrix, similar to the random-access memory in a computer. Like traditional computers, memory can be used to represent and execute a complex data structure, but, like a neural network, you can learn from the data. When conducting the supervised learning, we show that a DNC can successfully answer the problem of simulation, reasoning and argumentation in natural language. We show that he can learn the shortest distance and reasoning in a randomly generated diagram similar to a given point, and then generalize to a particular graph, for example: Transportation network and family tree structure. When the reinforcement learning, a DNC can complete the problem of moving block. In general, our results show that Dncs can solve complex, structured tasks, but these tasks are not external read-write memory.
Introduction :
Paper notes: Hybrid Computing using a neural network with dynamic external memory