6 Minutes
New mathematical model links memory capacity to sensory dimensions
Scientists at Skoltech have developed a mathematical model of memory that points to an unexpected conclusion: a conceptual space with seven dimensions may maximize the number of distinct memories an organism or artificial agent can hold. The work, published in Scientific Reports, reframes memory as an evolving set of abstract objects and identifies an optimal number of sensory-like features for storing and discriminating concepts.
Scientists have uncovered a mathematical link between memory, senses, and intelligence, revealing that seven may be the magic number.
What the model represents: engrams and conceptual space
The Skoltech team builds on a long tradition in cognitive science and mathematical neuroscience that treats memory traces, or engrams, as the basic units of stored information. An engram is modeled as a sparse assembly of neurons distributed across brain regions; its 'content' is an abstract object described by multiple features. In human terms, the mental representation of a banana contains visual, olfactory, gustatory, tactile and contextual attributes. Those sensory or feature channels define the dimensionality of a conceptual space in which all engrams reside and interact.
In the model, engrams evolve over time: they sharpen or blur depending on how often they are reactivated by sensory input. Repeated activation corresponds to learning and consolidation; lack of activation produces forgetting. The authors studied the steady-state distribution of engrams that emerges after many interactions with stimuli and analyzed how many distinct concepts that distribution can support as the number of feature dimensions varies.
Key discovery: capacity peaks at seven dimensions
The mathematical analysis yields a striking result: the number of reliably distinct engrams stored in the steady state reaches a maximum when each concept is characterized by seven independent features. Put differently, a seven-dimensional conceptual space optimizes memory capacity in the model — analogous to having seven senses rather than five. According to the researchers, this optimum is robust across a wide range of model assumptions about the statistics of stimuli and the geometry of conceptual representations.
Professor Nikolay Brilliantov of Skoltech AI, a study co-author, summarized the implication: 'Our analysis indicates that when concepts are encoded with seven characteristic features, the number of distinct items that can be retained rises to a maximum. This is a theoretical result that may inform how we think about sensory channels in artificial systems and, speculatively, about biological sensing.'
The team notes one important modeling caveat: clusters of similar engrams that sit around a common center are treated as a single concept for capacity calculations, which affects how 'distinct' memory items are counted.
Scientific context and methodology
The study follows mathematical approaches from early 20th-century theories of memory and modern statistical physics. Rather than running neural-network simulations alone, the researchers derived analytical expressions that describe how engram ensembles converge to a mature distribution. By scanning conceptual-space dimensionality and counting distinct attractor-like engram clusters at equilibrium, they identified the capacity peak at seven dimensions.
This analytical perspective complements experimental neuroscience and computational modeling by making clear predictions about how adding feature channels — for example, new types of sensors or modalities — should change an agent's memory capacity.
Implications for AI, robotics, and neuroscience
If the conclusion generalizes beyond the simplified model, it has practical implications:
- Robotics and AI: Designers of embodied agents and multimodal models may improve learning and recall by integrating a broader set of orthogonal sensors (e.g., magnetic, thermal, vibrotactile) that provide independent feature axes in representational space.
- Machine learning: Richer multimodal embeddings with carefully chosen independent channels could increase a model's capacity to store and disambiguate concepts without dramatically enlarging network size.
- Neuroscience and evolution: The result suggests hypotheses for comparative studies of sensory ecology and memory capacity across species, and it invites controlled experiments testing whether adding artificial modalities alters memory performance.
However, the authors emphasize that mapping model dimensions to biological 'senses' is speculative. Evolutionary, developmental and metabolic constraints all shape sensory systems, and real nervous systems organize inputs in highly correlated, hierarchical ways that differ from idealized independent features.
Future directions and experiments
The next steps are empirical tests and engineering trials. In neuroscience, researchers could examine whether augmenting sensory inputs (for instance, by providing people with a wearable magnetic field sensor) produces measurable changes in memory discrimination consistent with the model's predictions. In AI, controlled ablation and addition of orthogonal modalities in agents trained on the same tasks would test whether a seven-channel encoding architecture outperforms others in storing distinct concepts.
Expert Insight
Dr. Elena Park, cognitive AI researcher at the Global Cognitive Systems Lab, commented: 'The Skoltech result is a useful guidepost. It doesn't mean biology must have exactly seven senses, but it suggests that memory capacity is sensitive to the number and independence of representational axes. For engineers, the takeaway is practical: adding well-chosen, independent sensor channels can be more effective than simply scaling up model size.'
Conclusion
The Skoltech mathematical model offers a concise, testable prediction: a seven-dimensional conceptual space maximizes steady-state memory capacity under their assumptions. While the direct mapping from dimensions to biological senses remains speculative, the finding provides a fresh theoretical lens for research in cognitive neuroscience, artificial intelligence and robotics. Empirical studies and engineering prototypes that vary the number and independence of sensory channels will be essential to evaluate how this theoretical optimum translates into living brains and artificial agents.
Source: sciencedaily
Leave a Comment