Artificial intelligence Google DeepMind learns to use the extra memory
Combining complex mathematical data processing capabilities of artificial neural networks, the researchers units of Google DeepMind has created what can be considered one of the first implementations of a universal neuromorphic computers.
This hybrid smart car are able to find solutions to quite complex tasks, for example, to determine the best route in the metro, and not having a full and detailed diagrams.
When it comes to neuromorphic computers and artificial intelligence systems, commonly used artificial neural networks through the process of pre-training. However, these neural networks have, or rather had, one significant disadvantage – they are unable to dynamically allocate additional memory for data if required to solve their task.
The solution to the problem with dynamic memory allocation was the construction of the architecture of the so-called neuromorphic computer differentiable (differentiable neural computer, DNC). Essentially this computer is a neural network a new type associated with the external memory array. This computer also does not require any programming, it is enough to go through the process of preliminary training. And accumulated as a result of the learning experiences and knowledge are stored in the DNC-the computer not only in the structure of the neural network, but in external memory, which significantly extends its capabilities.
“After the training we have under our control, DNC-the computer was able to provide answers to complex synthetic questions that have been designed specifically to ensure that the computer was able to emulate the reasoning, to draw conclusions, to use the data and display the results, presenting them in natural language” – written by representatives of DeepMind, – “in Other words, our computer is thinking almost as a man.”
So what are the features of the new computer? According to available information, the machine have already been able to understand data complex family tree, some branches of which contained false information, and based on this analysis, the computer was able to answer some of the questions of the historical plan. When the computer was given the opportunity to familiarize yourself with the map of the London underground, he immediately was able to build the optimal multi-stage route to move from one station to another. Note that this problem is causing difficulties for the majority of normal people.
The team of specialists DeepMind is considering its product as a stand alone “artificial intelligence”, but as a kind of “multiplier” for human ingenuity. “Many challenges ranging from climate change analysis to the need for new methods of treatment of various diseases, suffer from their slow advance. Their complexity makes people unable to quick search for suitable solutions to these problems,” – wrote the representatives of DeepMind, “We hope that our artificial intelligence system in the future will be able to provide invaluable assistance in solving these and many other complex tasks.”
