Summary: Researchers have identified a notable parallel between how artificial intelligence (AI) models process memories and how the human hippocampus functions. This discovery links concepts from neuroscience and machine learning, revealing that memory consolidation—the transformation of short-term information into stable long-term representations—operates in strikingly similar ways in both Transformer-based AI models and biological brain tissue.
Focusing on the Transformer architecture, a central model class behind many modern AI advances, the interdisciplinary team found that the model’s internal memory operations resemble the gating mechanism of the NMDA receptor in the hippocampus. This insight not only informs efforts to build more efficient, brain-like AI but also offers a novel computational perspective for studying human memory processes.
Key Facts:
- The study identifies a significant similarity between memory processing in Transformer-based AI models and hippocampal function in the human brain.
- Researchers show the Transformer’s internal gating dynamics mirror the NMDA receptor’s role as a selective gate important for memory consolidation.
- These results point to opportunities for building AI that is both more memory-efficient and more interpretable, while also supplying a computational tool to explore biological memory mechanisms.
Source: Institute for Basic Science
An interdisciplinary team from the Center for Cognition and Sociality and the Data Science Group at the Institute for Basic Science (IBS) reports a clear resemblance between memory consolidation mechanisms in AI and the hippocampus of the brain.
Their work offers a fresh viewpoint on memory consolidation—the process that converts transient representations into long-term memories—by showing that the same functional principles can appear across biological and artificial systems. Understanding these shared principles is valuable both for advancing AI toward more general, human-like intelligence and for using AI models as tools to probe cognitive neuroscience questions.
Much of today’s progress toward Artificial General Intelligence (AGI) relies on scalable architectures and learning algorithms. Among these, the Transformer model has become foundational in natural language processing and other domains. The IBS team investigated whether core features of Transformer memory operations correspond to known neurobiological mechanisms that enable stable long-term storage in the brain.
At the center of the comparison is the NMDA receptor, a well-known molecular gate in the hippocampus that is essential for forming and stabilizing memories. In biological neurons, the NMDA receptor allows ions to pass only when specific conditions are met: the neurotransmitter glutamate must bind and a magnesium ion must be removed from the channel’s blocking position. This conditional opening controls when and how synaptic changes occur during learning.
In their analysis, the researchers observed that Transformer models implement a form of selective gating that plays a comparable role: it regulates when new information is integrated into the model’s internal memory representations and when existing representations are preserved. By drawing this parallel, the team probed whether manipulating the Transformer’s gating parameters could mimic the effects observed in biology.
One notable observation from neuroscience is that reduced magnesium levels can impair memory in animals by altering NMDA receptor function. Translating this idea, the researchers adjusted the Transformer’s gating dynamics and demonstrated that such changes can influence the model’s ability to retain long-term information. Specifically, designing the model’s update rules to emulate NMDA-like gating improved its long-term memory retention, suggesting that architectures inspired by biological nonlinearity can yield practical benefits.
This study thereby establishes a two-way dialogue: neuroscience helps inspire modifications to AI architectures that enhance memory behavior, and AI models provide a controlled setting to test hypotheses about how gated mechanisms support consolidation in the brain. The result is a stronger conceptual bridge between how brains and machines store, protect, and update information over time.
C. Justin LEE, neuroscientist and director at the institute, commented, “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 from KAIST added, “The human brain is remarkably energy efficient compared with the large AI models that require substantial computational resources. Our work opens new possibilities for low-cost, high-performance AI systems that learn and remember information more like humans do.”
What makes this work distinctive is its integration of brain-inspired nonlinearity into an AI framework, moving beyond superficial analogies to implement and test mechanisms that mirror biological gating. Beyond engineering gains, these experiments deepen our theoretical understanding of consolidation by framing it in computational terms that are directly testable within existing AI platforms.
The convergence of cognitive neuroscience and AI design promises two major benefits: the development of more efficient, interpretable AI systems that require less energy to learn and remember, and the provision of new computational models that neuroscientists can use to generate and evaluate hypotheses about human memory.
About this AGI and AI research news
Author: William Suh
Source: Institute for Basic Science
Contact: William Suh – Institute for Basic Science
Image: The image is credited to Neuroscience News