Human brains consolidate memories during sleep. Researchers are exploring whether AI systems could benefit from a similar approach. While some experts advocate for developing AI models that sleep and dream to enhance performance and reliability by emulating human brain functions, others caution against solely replicating human intelligence in AI research.
Concetto Spampinato and his team at the University of Catania, Italy, sought to address “catastrophic forgetting,” a phenomenon where AI models lose proficiency in previously mastered tasks when learning new ones. They introduced wake-sleep consolidated learning (WSCL), a novel training method inspired by how human brains solidify new information during sleep. This technique involves consolidating short-term memories into long-term storage, akin to humans processing and integrating daily experiences while asleep.
In WSCL, AI models undergo traditional training on a dataset during the “awake” phase and then transition into “sleep” periods. During these sleep phases, the models review recent data along with past learnings, preventing the loss of previously acquired skills. Additionally, the models experience a “dreaming” phase where they encounter entirely new data combinations, stimulating the formation of complex patterns and facilitating the assimilation of fresh knowledge.
Spampinato’s experiments demonstrated the efficacy of WSCL compared to conventional training methods, showing a notable improvement in accuracy and the retention of old tasks in AI models. Despite these positive outcomes, some experts caution against overly anthropomorphizing AI design or strictly imitating human brain structures. Andrew Rogoyski from the University of Surrey suggests that while the sleep training approach is intriguing, AI development should not be limited to mimicking human intelligence. He proposes considering alternative biological models like dolphins, which can rest one part of their brain while staying alert with another, as a source of inspiration for AI design.
In conclusion, while the concept of sleep-based learning for AI shows promise, the debate continues on the most effective path for advancing artificial intelligence beyond mere replication of human cognitive processes.
