In recent years, the advancement of artificial intelligence (AI) has enabled developers to create systems capable of emulating various human abilities, such as image recognition and question answering. However, unlike the human brain, which can deteriorate with age, AI systems typically maintain or even enhance their performance over time.
Researchers at the University of California, Irvine, have recently embarked on a study to mimic aging and biological neurodegeneration in AI agents. Their paper, which was pre-published on arXiv, has the potential to guide the development of innovative AI systems that utilize ‘artificial neurodegeneration’ to perform specific tasks.
Yu-Dai Tsai, co-author of the paper, shared that the inspiration for this study arose from discussions with Dr. Baldi and Dr. Pishgar, covering topics in neurodegeneration, learning, and AI safety. Additionally, Tsai’s personal experience with his father’s cognitive decline following brain trauma sparked new perspectives on the subject’s relevance to computer science, particularly in deep learning.
The researchers aimed to induce cognitive decline in AI agents to gain insights into complex systems, potentially improving their interpretability and security. They utilized IQ tests conducted by large language models (LLMs), specifically the LLaMA 2 model, to introduce the concept of ‘neural erosion.’ This deliberate erosion involved removing synapses or neurons or adding Gaussian noise during or after training, leading to a controlled decline in the LLMs’ performance.
Their findings revealed a distinct pattern of decline in AI systems, mirroring neurodegeneration observed in humans. As artificial synapses and neurons were removed, the AI systems first lost abstract thinking abilities, followed by mathematical degradation, and ultimately, linguistic skills. This pattern aligned with neurodegeneration patterns observed in humans.
Tsai emphasized that this study marks the beginning of a series, with plans to develop specific tests for AI systems and extend the emulation to other neural diseases and neurodiversity. Furthermore, they aim to apply their methods to enhance AI security and interpretability. While collaborations with neuroscientists are welcomed, their primary focus remains on exploring new frontiers in AI studies, rather than replicating human brain diseases.
Source: techxplore.com
