Remember meForgot password?
    Log in with Twitter

article imageLetting an AI go 'to sleep' enhances its performance

By Tim Sandle     Feb 15, 2019 in Technology
An artificial intelligence has been programmed so that it enters into 'sleep mode' with the aim of enhancing performance. The process is based on a mathematical model.
The type of artificial intelligence being piloted for 'sleep mode' is an artificial neural network, which is based on biological neural networks. Such systems would never ordinarily switch off and 'sleep'. However, to simulate this very human activity and to encourage background activity to continue (inducing something analogous to 'dreaming'), Italian scientists have programmed a type of artificial neural network termed a Hopfield network to do just this.
A Hopfield network is a type of recurrent artificial neural network, where the system functions as a content-addressable ("associative") memory system with binary threshold nodes. Hopfield networks have been used to provide a working model for understanding human memory.
The basis of the research was to determine whether 'sleep' benefits an artificial intelligence system the same way as it benefits people.
Sleep is essential for human brain health. The sleep process enables neurons to address all the unnecessary synaptic connections that are formed during the day. The physiological term for this is synaptic homeostasis, and it is necessary for preventing the brain from being overrun by useless memories. This is one of the reasons why a good night's sleep is necessary for maintaining cognitive performance.
The research suggests that something similar could be occurring when artificial neural networks are allowed to enter 'sleep mode' and begin to 'dream.' The researchers defined an “awake regime” and a “sleep regime” for their associative neural network. When they forced the network “to sleep”, they assessed the extent that the clearness of spurious stored patterns and consolidation of pure ones occurred.
The results were very successful, according to Science Alert. Without sleep, the maximal capacity of the network was rated α=0.14 (the α symbol represents the number of stored bits per synapse); however, when a sleep cycle was incorporated, the network reached its theoretical limit - α=1. Hence, the findings suggest that sleeping is as essential for an artificial intelligence, as it is for a biological one.
The research has been published in the journal Neural Networks, in a paper titled "Dreaming neural networks: Forgetting spurious memories and reinforcing pure ones."
More about Artificial intelligence, machine learning, Intelligence, Computers
More news from
Latest News
Top News