Researchers Find Striking Similarity Between AI and Human Memory Processes
Researchers at the Institute for Basic Science have made a groundbreaking discovery, revealing a remarkable parallel between artificial intelligence (AI) memory processing and the human brain’s hippocampal functions. The study, conducted by an interdisciplinary team from the Center for Cognition and Sociality and the Data Science Group, delved into the memory consolidation process of AI models, shedding new light on how these systems transform short-term memories into long-term ones.
The focus of the research was on the Transformer model, a critical component of AI advancements. By examining the model’s memory processes, the researchers uncovered intriguing similarities to the NMDA receptor mechanism in the human brain. This finding not only pushes forward the development of Artificial General Intelligence (AGI) but also deepens our understanding of the intricacies of human memory systems.
Memory consolidation plays a vital role in the quest for AGI, with organizations like OpenAI and Google DeepMind leading the charge. The Transformer model, which lies at the heart of these efforts, is now under a new lens of exploration.
To comprehend how powerful AI systems learn and retain information, the research team turned to the principles of human brain learning, particularly focusing on memory consolidation through the NMDA receptor located in the hippocampus.
The NMDA receptor acts as an intelligent gateway in the brain, facilitating learning and memory formation. The presence of a brain chemical called glutamate triggers the excitation of nerve cells while a magnesium ion functions as a gatekeeper, blocking the gate. Only when this gatekeeper is removed, can substances flow into the cell, enabling the creation and retention of memories.
What the researchers discovered was truly remarkable. The Transformer model appeared to employ a gating process akin to the brain’s NMDA receptor. This revelation prompted the team to explore whether the Transformer’s memory consolidation could be influenced by a mechanism similar to the NMDA receptor’s gating process.
In animal brains, low magnesium levels are known to weaken memory function. Intriguingly, the researchers found that by mimicking the NMDA receptor, they could enhance the long-term memory capabilities of the Transformer model. Similar to the brain, where altering magnesium levels impacts memory strength, making adjustments to the Transformer’s parameters to align with the gating action of the NMDA receptor resulted in improved memory performance in the AI model.
This groundbreaking discovery suggests that AI models’ learning processes can be explained by established neuroscience knowledge. By bridging the gap between AI and neuroscience, researchers can delve deeper into the operational principles of the human brain and develop more advanced AI systems.
C. Justin LEE, a neuroscientist director at the institute, expressed his excitement about this research, stating, This research marks a crucial step in advancing both AI and neuroscience. It enables us to gain deeper insights into the brain’s functioning and develop more sophisticated AI systems based on these findings.
CHA Meeyoung, a data scientist involved in the study, emphasized the potential for low-cost, high-performance AI systems that mimic human learning capabilities. She highlighted how the human brain operates with minimal energy compared to the resource-intensive AI models, making their work a gateway to unlocking the possibilities of more efficient AI systems.
What sets this study apart is the incorporation of brain-inspired nonlinearity into the design of AI systems, which signals a significant breakthrough in emulating human-like memory consolidation.
The convergence of human cognitive mechanisms and AI design not only holds promise for creating low-cost, high-performance AI systems but also provides invaluable insights into the workings of the brain through AI models.
This cutting-edge research presents a remarkable avenue for both AI and neuroscience, propelling our understanding of memory processes and paving the way for the development of more advanced, efficient, and human-like AI systems.
As we progress in unraveling the mysteries of the mind, the possibilities for transformative advancements in technology are vast. The integration of AI and neuroscience brings us one step closer to unlocking the true potential of artificial intelligence.
