In an era where the quest for Artificial General Intelligence (AGI) is picking up pace, remarkable insights from the Institute for Basic Science (IBS) have been unveiled, showcasing a striking similarity between the memory processing of artificial intelligence (AI) systems and the human brain. This parallel between how AI models memory-forming mechanism functions and the hippocampus’s role in human memory provides a transformative perspective on memory consolidation, paving the way for low-cost, high-performance AI systems.
The AI Transformer Model and Hippocampus
At the heart of these advancements is the Transformer model, a cutting-edge AI that has become central to efforts by influential entities like OpenAI and Google DeepMind to replicate human-like intelligence. Researchers from the Center for Cognition and Sociality and the Data Science Group within IBS have discovered that the Transformer employs a gatekeeping process akin to the NMDA receptor in the hippocampus—crucial for converting short-term memories into long-term ones.
The NMDA receptor functions as a ‘smart door’ in the brain, with glutamate, a brain chemical, acting as a key that triggers a nerve cell’s excitation. Meanwhile, a magnesium ion plays a role similar to a security guard, enabling substances to flow into the cell only when it steps aside, a process essential for creating and retaining memories.
By mimicking this gatekeeping process, researchers found that long-term memory in the Transformer model could be enhanced, just as changing magnesium levels can affect memory strength in the animal brain. Adjusting the Transformer’s parameters to reflect this gating action led to a substantial improvement in AI memory function—akin to the efficient energy management observed in the human brain system.
Mimicking Human Brain for Advanced AI Models
Neuroscientist director C. Justin Lee remarked, “This research makes a crucial step in advancing AI and neuroscience. It allows us to delve deeper into the brain’s operating principles and develop more advanced AI systems based on these insights.” Data scientist CHA Meeyoung added, “Our work opens up new possibilities for AI systems that learn and remember information like humans, operating with minimal energy, unlike current large AI models that need immense resources.”
This integration of brain-inspired nonlinearity into AI models is not just a significant advancement in simulating human-like memory consolidation but also enriches our understanding of cognitive mechanisms within the human brain. It sets a precedent for continued convergence between the study of the human brain and the evolvement of artificial intelligence.
A Step Closer to Artificial General Intelligence
The new methodology provides a compelling narrative that links the distinct systems of the brain’s hippocampus, renowned for its storage of long-term memories and spatial navigation, with the technological forefront of AI. By crafting AI that can process memories similarly to the way we do, a new horizon in artificial intelligence emerges—bringing us one step closer to AGI while offering invaluable insights into the intricacies of our own consciousness and memory formation.
The findings from IBS suggest a future where AI can be both high-performing and sustainable, requiring lower resources to operate much like the human brain. The development of such systems could potentially change the landscape of AI, yielding machines capable of intricate thought and recollection processes reflective of the nonlinearity and adaptability found in the most complex organism known—the human brain.
