Nobody can tell whether androids would dream of electric goats, but they will almost certainly need rest periods that provide benefits close to those that give living brains sleep, according to new Los Alamos National Laboratory research, it is clear that they are researching spiking neural networks which are of structures that would help to learn a lot of things like living brains which is told by Yijing Watkins, a computer scientist who is at Los Alamos National Laboratory. “We were intrigued by the possibility that a neuromorphic processor could be trained in a way close to how humans and other biological systems learn from their environment during their childhood growth. Watkins along with her research team observed that the miniature of the network became unstable even after the continuous-time of unsupervised learning. Stability was restored when they exposed the networks to states analogous to the waves experienced by living brains during sleep.” It was like we gave the neural networks the equivalent of a good night’s rest, “Watkins said.
The discovery came about when the research team was working to develop neural networks that closely approximate how people and other biological systems learn to see.
Garrett Kenyon, a computer scientist and the co-author of Los Almost said that when we are trying to use biologically realistic, spearing neuromorphic processors or trying to discover the biology itself, the question arises of how to keep learning systems more stable and Los Alamos study co-author. In the primary stage, the group found it really tough to stabilize the simulated neural networks which is undergoing unsupervised dictionary training that involves classifying objects without having previous examples to compare with them.
The researchers characterize the decision to expose the networks to an artificial sleep analogue as almost a last-ditch effort to stabilize them. They experimented with different types of noise, which is roughly comparable to the static that you might find between stations while tuning a radio. The best results came from using so-called Gaussian noise waves, which have a wide variety of frequencies and amplitudes. They assume the noise emulates the feedback provided by biological neurons during slow-wave sleep. The results indicate that slow-wave sleep can work, in part, to keep cortical neurons stable and not hallucinating.
The next aim of the groups is to implement their algorithm on the neuromorphic chip from Intel’s Loihi. They hope that allowing Loihi to sleep from time to time will allow it to process information stably in real-time from a silicone retina camera. If the results reflect the need for sleep in artificial brains, we should possibly expect the same to be true of potential androids and other smart machines.
