An artificial intelligence that mimics the operation of a standard computer within its neural network could speed up certain calculations. Researchers have used it to put an AI inside an AI and play Pong.
If you want a computer to do something, you have to write code that manipulates bits of data. But if you code an AI-driven neural network, you have to train it with feedback before it can do anything. For instance, a neural network can distinguish between photos of cats and dogs after being shown thousands of examples and told if its guesses are correct.
Now and at the University of Pennsylvania have a new approach, in which a neural network runs a code, just like an ordinary computer.
Advertisement
A neural network is a series of nodes, or artificial neurons, that takes an input and returns a changed output. The pair calculated what effect individual artificial neurons had, and used this to piece together a very simple neural network that could carry out basic tasks, such as addition.
Kim and Bassett then linked several networks together in chains so they could do more complex operations, replicating the behaviour of the logic gates found in computer chips. These chains were combined to make a network that could do things a classical computer can, including running a virtual neural network and a playable version of the game Pong.
The virtual neural network can also drastically simplify splitting up huge computational tasks. These are often spread over many processors to gain speed, but take more power to be split into chunks that can be run independently by separate chips, then recombined.
An emerging breed of machine called a neuromorphic computer, designed to efficiently run AI software, may also be able to help these virtual networks work faster. While a computer uses its CPU to carry out tasks, and stores data in memory, a neuromorphic computer uses artificial neurons to both store and compute, lowering the number of operations it must carry out. Neuromorphic computers may also make it easier for software to accurately work with continuous variables, such as those in physics simulations.
at Leipzig University in Germany says neural networks running code could squeeze better performance out of neuromorphic chips. 鈥淭here will definitely be specific applications where these computers are going to be outperforming standard computers. And by far, I believe.鈥�
at Montreal Polytechnic in Canada says this could be exciting, so long as the chains can avoid long calculations failing if an algorithm forgets the beginning as it is learning the end.
These neural networks would also need to be scaled up. 鈥淐omputers don鈥檛 just have one or two logic gates 鈥� a CPU will have billions of transistors,鈥� says Zayed. 鈥淛ust because it worked for two or three gates, that doesn鈥檛 necessarily mean that it will scale up to billions.鈥�
Reference:
