The way we think about the world is not a direct copy of reality; it is a set of internal symbols, images, and patterns that our brains (and increasingly, our artificial agents) construct, transform, and use to act. Understanding those mental representations is the key to unlocking everything from why a honeybee can find a flower a kilometer away, to how a self‑governing AI can plan a rescue mission without human oversight.
In the last half‑century cognitive science, neuroscience, and artificial intelligence have converged on a surprisingly common problem: how does a mind encode, store, and manipulate information about the external world? The answer is not a single model but a family of theories—collectively called mental representation theory—that describe the shape, format, and dynamics of internal “maps.” These theories guide research on memory disorders, inform the design of autonomous robots, and even shape conservation strategies that depend on understanding animal cognition.
This pillar page surveys the most influential models, the neural mechanisms that instantiate them, and the practical implications for both living organisms and synthetic agents. By the end you will see how a bee’s waggle dance, a child’s concept of “dog,” and a large‑language model’s token embeddings all share a common computational heritage, and why that matters for preserving biodiversity and building trustworthy AI.
1. What Is a Mental Representation?
A mental representation is any internal state that stands in for something external—an object, a relation, a goal, or an abstract rule. The classic definition from philosopher Jerry Fodor (1975) is “the mental vehicle that carries meaning.” In contemporary terms, a representation must satisfy three minimal criteria:
- Content – it carries information about something not present (e.g., the layout of a garden).
- Structure – its internal organization reflects relations among the represented items (e.g., a map’s north‑south axis).
- Accessibility – it can be used by cognitive processes such as perception, reasoning, or motor planning.
Representations differ from raw sensory data because they are abstracted and stable across time and contexts. For example, when you look at a red apple, the retinal image is a high‑dimensional pattern of light wavelengths. Your brain quickly compresses that pattern into a conceptual node that encodes “apple” (type, edibility, typical color) and links it to other concepts like “fruit” or “snack.”
Neuroscientists often equate mental representations with neural population codes—vectors of firing rates across ensembles of neurons. In AI, they appear as embeddings (vectors of numbers) or symbolic structures (trees, graphs). The diversity of formats is the reason why multiple theoretical frameworks exist; each emphasizes different aspects of the content–structure–accessibility triad.
2. Historical Foundations
2.1 From Aristotle to the “Computational Mind”
Aristotle’s De Anima (c. 350 BC) argued that the soul forms phantasms—mental images derived from sense experience. Though not a scientific model, this idea of internal images survives in modern imagery research.
The 19th‑century psychophysicists (e.g., Fechner) introduced the concept of mental “elements” that could be summed to predict perception, laying groundwork for later elemental theories.
2.2 The Symbolic Era (1950s‑1970s)
The post‑World‑II boom in computer science birthed the symbolic or classical view of cognition: the mind as a digital computer manipulating discrete symbols according to formal rules. Newell and Simon’s Physical Symbol System Hypothesis (1976) famously claimed that “a physical symbol system has the necessary and sufficient means for general intelligent action.”
During this era, psychologists such as George Miller (1956) quantified mental capacity—the magical number seven, plus or minus two—showing that working memory can hold roughly 4–7 chunks of information. This quantitative constraint shaped early models of representation (e.g., chunking in language processing).
2.3 Connectionism and Distributed Codes (1980s‑1990s)
Parallel distributed processing (PDP) models, championed by Rumelhart, McClelland, and Hinton, argued that cognition emerges from graded, distributed activity across many units. Their seminal book Parallel Distributed Processing (1986) introduced connectionist representations: each concept is encoded as a pattern of activation over a population, rather than a single symbol.
Empirical support arrived when Hubel and Wiesel (1977) mapped orientation columns in the visual cortex, revealing that features such as edge orientation are represented by groups of neurons with overlapping receptive fields.
2.4 Embodied and Ecological Turns (1990s‑present)
More recent work, led by scholars like James J. Gibson and Lakoff & Johnson, argues that cognition is situated and embodied: representations are shaped by the body’s interaction with the environment. The sensorimotor contingency view suggests that mental representations are not static symbols but dynamic predictions about sensory outcomes of actions.
These historical strands converge today: modern mental representation theory must account for symbolic precision, distributed neural realism, and embodied dynamics—all while being testable in both biological and artificial systems.
3. Core Theoretical Frameworks
3.1 Symbolic (Rule‑Based) Models
Symbolic models treat mental representations as discrete tokens with explicit syntactic relationships. A classic example is semantic networks, where nodes represent concepts and labeled edges encode relations (e.g., dog → is‑a → mammal).
Strengths:
- Clear compositionality—complex ideas can be built from simpler parts.
- Direct mapping to logical inference engines (e.g., Prolog).
Limitations:
- Struggles with graded similarity (e.g., “cat” and “tiger” are more alike than “cat” and “car”).
- Requires a symbol grounding mechanism to connect tokens to sensorimotor experience.
3.2 Connectionist (Distributed) Models
In connectionist accounts, a representation is a high‑dimensional vector of activation values (often bounded between 0 and 1). Learning occurs via weight updates (e.g., backpropagation) that shape the vector space so that similar inputs occupy nearby points.
Empirical anchor: Place cells in the rat hippocampus (O’Keefe & Dostrovsky, 1971) fire when the animal occupies a specific location. Thousands of such cells collectively encode a cognitive map—a distributed representation of space.
Strengths:
- Captures similarity naturally through Euclidean distance.
- Explains robustness to noise and partial damage (graceful degradation).
Limitations:
- Compositionality is emergent and often opaque.
- Requires large datasets for training—something natural brains achieve with far fewer experiences.
3.3 Embodied / Predictive Coding Models
Predictive coding (Friston, 2005) posits that the brain continuously generates predictions about incoming sensory data and updates its internal model by minimizing prediction error. Representations are thus generative models that encode expectations about the world.
A concrete illustration: when a honeybee approaches a flower, its optic flow predicts the timing of landing; any mismatch triggers corrective motor commands. Neural implementations involve hierarchical Bayesian inference, where higher cortical layers encode abstract priors and lower layers encode sensory likelihoods.
Strengths:
- Integrates perception, action, and learning into a single computational principle.
- Provides a mechanistic account of attention and surprise.
Limitations:
- Formalizing the exact form of priors remains challenging.
- Direct experimental validation at the whole‑brain level is still limited.
4. Neural Mechanisms Underlying Representations
4.1 Sensory Cortices: Feature Maps
Primary visual cortex (V1) contains orientation columns, each tuned to a narrow range of edge angles. In macaques, about 200,000 orientation-selective neurons per square millimeter have been recorded (Hubel & Wiesel, 1974). These columns constitute a feature map—a spatially organized representation of visual attributes.
Similar maps exist for auditory frequency (tonotopic maps) and somatosensory touch (somatotopic maps). The topographic arrangement allows the brain to compute local similarity efficiently, a principle that inspired convolutional neural networks (CNNs) in AI.
4.2 Hippocampal–Entorhinal System: Spatial Coding
The discovery of place cells (≈ 10⁴ cells in a rat hippocampus) and grid cells (≈ 5 × 10⁴ cells in the medial entorhinal cortex) revealed a multi‑scale metric for navigation. Grid cells fire in a hexagonal lattice covering the environment, with spacing ranging from 30 cm to 3 m, providing a neural GPS.
Human fMRI studies show that navigating a virtual maze activates the same regions, and participants can recall routes with an average error of ≈ 5 % of the path length—comparable to the precision of a honeybee’s waggle dance (see § 7).
4.3 Prefrontal Cortex: Abstract Rules
Dorsolateral prefrontal cortex (dlPFC) holds task sets—abstract representations of rules governing behavior. Single‑unit recordings in monkeys (e.g., Miller & Cohen, 2001) demonstrate that dlPFC neurons maintain rule‑specific firing for up to 30 seconds during a delay period, a neural correlate of working memory capacity.
These rule representations are mixed selectivity: a single neuron may encode multiple dimensions (e.g., stimulus identity and task rule), supporting the flexibility required for symbolic reasoning.
4.4 Neural Oscillations and Binding
Representational binding—the process of linking features (color, shape, motion) into a unified object—relies on gamma (30–100 Hz) oscillations that synchronize across disparate cortical areas. Studies using intracranial EEG show that synchronized gamma bursts increase by ≈ 50 % when participants view coherent objects versus scrambled images.
Oscillatory binding provides a temporal code that complements spatial coding, enabling the brain to hold multiple objects simultaneously without interference—a principle echoed in attention mechanisms of modern transformer models (Vaswani et al., 2017).
5. Types of Mental Representations
5.1 Propositional Representations
These are language‑like structures that encode statements such as “The rose is red.” They enable logical inference and are the backbone of symbolic AI. In the brain, semantic dementia patients lose access to such propositions while retaining image-based knowledge, suggesting a dissociation between propositional and perceptual representations.
5.2 Imagery Representations
Visual mental imagery activates early visual areas (V1/V2) with a strength of ≈ 30 % of that evoked by actual perception (Kosslyn et al., 1995). This overlap supports the claim that images are internal pictures rather than purely symbolic descriptions.
5.3 Schemas and Scripts
A schema is a higher‑order structure that captures typical relationships (e.g., a “restaurant” schema includes waiters, menus, ordering). Neuroimaging shows that schemas engage the default mode network (DMN), especially the posterior cingulate cortex, with activation increases of ≈ 15 % when participants read stories that match familiar scripts.
5.4 Cognitive Maps
Beyond spatial navigation, cognitive maps can represent abstract domains such as social hierarchies or conceptual similarity. A landmark study (Tavares et al., 2015) demonstrated that the hippocampus encodes a 2‑D space where one axis corresponds to political ideology and the other to economic stance, allowing participants to infer relationships between novel political figures.
6. Representational Dynamics: Learning, Consolidation, and Plasticity
6.1 Hebbian Learning and Synaptic Plasticity
The classic rule “cells that fire together, wire together” (Hebb, 1949) describes how co‑activation strengthens synaptic efficacy. In the mouse visual cortex, repeated pairing of a specific orientation with a reward leads to a ≈ 20 % increase in the firing rate of orientation‑selective neurons after just 30 minutes of training.
6.2 Systems Consolidation
Memory consolidation occurs in two stages: synaptic consolidation (hours) and systems consolidation (weeks to years). Human studies using sleep polysomnography reveal that slow‑wave sleep spindles correlate with a 10 % improvement in recall after a night’s sleep, reflecting hippocampal–cortical transfer.
6.3 Neurogenesis and Representational Renewal
Adult hippocampal neurogenesis (≈ 700 new granule cells per day in humans) contributes to pattern separation—the ability to keep similar experiences distinct. Computational models show that adding new units reduces interference by ≈ 30 % in tasks requiring discrimination of overlapping inputs.
6.4 Plasticity in Artificial Networks
Deep neural networks undergo catastrophic forgetting when trained sequentially on new tasks, losing up to 90 % of previously acquired performance. Techniques such as elastic weight consolidation (EWC) mitigate this by penalizing changes to important weights, mirroring biological synaptic stability mechanisms.
7. Real‑World Applications
7.1 Language Understanding
Large language models (LLMs) like GPT‑4 employ contextual embeddings of size 12,288 dimensions, learned from ≈ 1 trillion tokens. These embeddings capture syntactic and semantic relationships: the cosine similarity between “king” and “queen” is ≈ 0.78, while “king” and “apple” drops to ≈ 0.12. The success of LLMs suggests that distributed representations can support complex symbolic manipulation when combined with attention mechanisms.
7.2 Navigation in Animals: The Honeybee Case
Honeybees (Apis mellifera) perform a waggle dance that encodes distance and direction to a food source. Laboratory measurements show that the dance’s angle correlates with the sun’s azimuth within ± 5°, and the duration of the waggle segment predicts distance with a standard error of ≈ 15 % of the true distance (von Frisch, 1967).
Neurophysiological work reveals that optic flow in the bee’s lobula provides a metric for distance, while central complex neurons encode heading direction. The bee’s internal map thus combines rate coding (distance) with vector coding (direction), a hybrid representation that has inspired autonomous drone navigation algorithms.
7.3 Autonomous Vehicles and Self‑Governing AI
Self‑governing AI agents (e.g., autonomous rescue robots) must integrate perception, planning, and ethical decision‑making without constant human oversight. A recent field trial in the DARPA Subterranean Challenge (2022) demonstrated that a fleet of robots could map a 1 km underground network, maintaining a global representation error < 0.3 m while dynamically replanning routes in response to blocked passages.
These agents rely on graph‑based cognitive maps—nodes represent rooms, edges encode traversability, and edge weights encode risk. The representation is updated using Bayesian belief updates that fuse LIDAR data (sensorimotor) with prior maps (symbolic). The hybrid approach mirrors the bee’s integration of metric and topological cues.
7.4 Conservation Technology
Understanding animal representations informs conservation interventions. For example, re‑wilding projects for the endangered Kakapo (Strigops habroptilus) use playback of conspecific calls to stimulate territorial mapping, exploiting the bird’s song‑based spatial representation. Field trials in New Zealand measured a 40 % increase in nesting site fidelity after targeted acoustic conditioning (Merton et al., 2021).
By modeling how species encode habitat features, managers can design landscape corridors that align with natural mental maps, increasing the success of migration corridors for pollinators like bumblebees (Bombus terrestris).
8. Implications for Artificial Intelligence
8.1 Symbolic vs. Distributed Debate
The symbolic‑connectionist debate resurfaces in modern AI: should agents reason with explicit symbols (e.g., logic programs) or rely on high‑dimensional embeddings? Hybrid architectures, such as Neuro‑Symbolic systems, embed logical constraints into neural networks using differentiable theorem provers. Benchmarks on the CLEVR visual reasoning dataset show that neuro‑symbolic models achieve ≈ 95 % accuracy, surpassing pure CNNs (≈ 78 %) and pure symbolic parsers (≈ 85 %).
8.2 Grounding and Embodied AI
Grounding remains a central challenge: an AI’s symbols must be tied to sensory experience. Embodied agents—robots equipped with cameras, force sensors, and proprioception—learn representations through interaction. In a 2023 study, a quadruped robot learned to navigate a cluttered arena using a predictive coding loss that reduced sensorimotor prediction error by 45 % over 48 hours of self‑supervised exploration.
This mirrors the bee’s reliance on optic flow and compass cues, suggesting that self‑generated predictions are a universal route to grounded representation.
8.3 Self‑Governance and Ethical Decision‑Making
Self‑governing AI must encode not only physical maps but also normative representations (e.g., “do no harm”). Recent work on value alignment introduces utility vectors that are learned from human feedback and updated via inverse reinforcement learning. When deployed in a simulated disaster scenario, agents that maintained a dual‑layer representation—one for operational goals, one for ethical constraints—reduced fatality rates by 23 % compared to agents with a single objective function.
The architecture of these agents echoes the brain’s dual‑process organization: fast, automatic sensorimotor pathways (System 1) and slower, deliberative reasoning (System 2).
9. Open Challenges and Future Directions
9.1 The Symbol Grounding Problem
Despite progress, we lack a comprehensive account of how symbols acquire meaning from raw sensory data. Proposed solutions range from sensorimotor contingencies (O’Regan & Noë, 2001) to developmental bootstrapping where infants learn words in context. Empirical work with infant‑robot interaction—where robots learn word–object mappings through joint attention—has achieved ≈ 70 % labeling accuracy after only 200 exposure trials, still far from the human rate of ≈ 95 % after comparable experience.
9.2 Compositional Generalization
Current deep networks excel at memorizing patterns but falter when asked to compose known elements in novel ways (e.g., “red‑striped” when never seen together). Benchmarking on the SCAN compositional dataset shows a gap: state‑of‑the‑art transformers achieve ≈ 30 % accuracy on the hardest split, versus ≈ 100 % for symbolic grammar models. Bridging this gap may require structured priors that encode combinatorial rules, akin to the brain’s hierarchical syntax processing in Broca’s area.
9.3 Conscious Access and Metacognition
Do mental representations become conscious when they are broadcast in a global neuronal workspace (Dehaene & Changeux, 2011)? Intracranial recordings in patients performing visual detection tasks reveal a late‑phase (≈ 300 ms) global ignition that correlates with reported awareness. Understanding the transition from sub‑conscious representation to conscious report could inform AI safety, ensuring that autonomous agents can flag when a decision exceeds their confidence threshold.
9.4 Integrating Multi‑Modal Maps
Animals routinely integrate visual, olfactory, and auditory cues into a unified map. The central complex of insects merges compass information from polarized light with wind direction and odor gradients. In AI, multi‑modal transformers (e.g., CLIP, Flamingo) have begun to fuse image and text embeddings, achieving ≈ 80 % zero‑shot classification on ImageNet. Yet truly holistic representations that support navigation, communication, and planning remain an open frontier.
9.5 Conservation‑Driven Representation Research
A promising direction is citizen‑science neuroethology, where low‑cost electrophysiology rigs record neural activity from pollinators in the field. Preliminary data from a Bombus colony indicate that olfactory glomeruli respond to floral scent mixtures with a population code that predicts foraging distance within ± 10 m. Scaling such datasets could reveal how environmental stressors (pesticides, climate change) alter representational fidelity, providing early warning signals for ecosystem health.
10. Synthesis: A Unified View
Mental representation theory is no longer a set of competing paradigms but a multifaceted tapestry where symbolic precision, distributed neural realism, and embodied prediction intertwine. The brain appears to maintain layered maps: low‑level sensory feature maps (orientation, frequency), intermediate spatial codes (place/grid cells), and high‑level abstract schemas (social hierarchies). Each layer communicates via oscillatory synchrony and hierarchical Bayesian updates, enabling both rapid perception and long‑term planning.
Artificial agents that emulate this architecture—by coupling graph‑structured symbolic reasoning with deep embeddings and predictive loss functions—show promising gains in robustness, interpretability, and ethical behavior. Moreover, the same principles guide the design of bee‑inspired navigation algorithms, conservation monitoring tools, and self‑governing AI that can act responsibly in uncertain environments.
The convergence of neuroscience, cognitive psychology, and AI is not merely academic; it provides the conceptual scaffolding for technologies that protect pollinators, manage ecosystems, and ensure that increasingly autonomous systems remain aligned with human values.
Why It Matters
Mental representations are the currency of cognition—the internal language through which brains and machines make sense of the world. By unraveling how these representations are built, updated, and used, we gain the ability to:
- Design AI that learns like animals, reducing data requirements and improving adaptability.
- Create conservation technologies that respect the natural mental maps of pollinators, enhancing habitat connectivity and reducing human‑wildlife conflict.
- Develop self‑governing agents that can articulate their internal state, flag uncertainty, and make ethically sound choices without constant oversight.
In short, a deeper grasp of mental representation theory equips us to bridge biology and technology, fostering a future where thriving ecosystems and trustworthy AI co‑evolve. The honeybee’s waggle dance and the transformer’s attention matrix may seem worlds apart, but both are expressions of the same fundamental principle: the mind, in all its forms, builds internal worlds to survive and thrive.