A recent study has shown that AI brains might need a good night's sleep to ensure reliable performance, similar to the human brain. Research in AI has been progressing at a brisk pace and there have been quite a few remarkable breakthroughs highlighting what artificial intelligence can now do. While the performance seems to be going in the right direction, there are still some issues that need to be worked out to ensure AI systems are safe for practical use.
AI is one of the emerging technologies that's expected to have a significant impact on life and society. The reliance on AI has become even clearer during the ongoing COVID-19 pandemic with AI tools having been used in various ways to help manage the public health crisis. For example, a company funded by the US government is using machine learning software to fight coronavirus misinformation spreading online, while other use-cases have shown how AI could be beneficial in diagnosing medical conditions. Although there are a number of ways AI computing systems are being used, neural networks have been greatly inspired by the development of animal brains.
A recent Los Almos National Laboratory study suggests that neural network systems can benefit from periodic rests, just like human brains. The research team, headed by computer scientist Yijing Watkins, made the discovery while studying neural networks that process information similar to how human brains do. The research was specifically interested in developing a neural network that closely resembles how biological systems learn by observing their environment.
A Good Night's Sleep Could Be Beneficial To AI
The researchers were struggling to stabilize the neural networks while they were undergoing training to classify objects without prior samples to compare them to. This led to the observation that AI network simulations become unstable after long periods of continuous learning. The research team tried to solve this issue by exposing the networks to different kind of noise - similar to the static humans hear while tuning to different radio stations. However, it was the artificial use of waves that a biological brain experiences while sleeping that eventually worked. In a bid to test the suggestion further, the researchers are preparing to test their findings on Intel's Loihi neuromorphic processor. If Loihi is able to process information more reliably after occasional bouts of sleep, then a similar approach could be used with other AI systems, including androids. For reference, sleep-induced stability is only applicable to neuromorphic processors that attempt to mimic the biological brain.
With AI predicted to change society radically in the future, it is important for researchers to better understand this technology and identify ways that can help to reduce errors as much as possible. This is especially true in cases where humans are expected to trust an automated artificially intelligent system to manage complex tasks with a tangible risk factor, including running a fleet of self-driving cars or diagnosing medical conditions.
Source: Los Almos National Laboratory