Study Shows Diverse Neural Networks Perform Better, Thanks to AI’s Ability for Self-Tuning
An artificial intelligence (AI) system that can fine-tune its own neural network by embracing diversity has been found to perform better, particularly in solving complex tasks, according to a new study carried out by researchers at North Carolina State University.
The Significance of Diversity in AI
The study, led by William Ditto, professor of physics and director of NC State’s Nonlinear Artificial Intelligence Laboratory, aimed to understand whether AI would choose diversity over lack of diversity and how that choice would affect its performance.
Neural networks, which are a type of advanced AI based on how our brains work, rely on strong connections between artificial neurons. In the conventional approach, these networks are made up of large numbers of identical neurons with fixed connections. In contrast, the team gave their AI the ability to choose the number, shape, and connection strength between neurons, leading to the creation of diverse sub-networks within the neural network.
The Benefits of AI’s Ability for Self-Tuning
This meta-learning approach allowed the AI to modify the composition of its neural network, optimizing its performance. By changing the type and mixture of artificial neurons, the AI could solve a problem, examine the result, and improve its performance. The study found that the AI consistently chose diversity over homogeneity as a way to strengthen its performance.
Improved Accuracy and Performance
To test the accuracy and performance of the AI, the researchers conducted a numerical classifying exercise. The diverse AI scored an accuracy rate of 70%, compared to only 57% accuracy with a standard, homogenous AI.
Furthermore, the researchers discovered that the diversity-based AI was up to 10 times more accurate in tackling complex problems, such as predicting pendulum swings or galactic motion, compared to conventional AI.
The Implications of the Study
This study highlights the importance of diversity within neural networks for enhancing AI performance. By giving AI the ability to look inward and learn how it learns, researchers have unlocked the potential for AI to adapt and optimize its own neural structure. This self-tuning ability has the potential to significantly improve the efficiency and accuracy of AI in solving complex problems.
The research article, titled “Artificial Neural Networks Able to Embrace Diversity Improve Ability to Solve Complex Problems,” was published in Scientific Reports. The study was supported by the Office of Naval Research and United Therapeutics, with contributions from researchers at the Indian Institute of Science Education and Research Mohali.