Mental Representation: From Philosophy to Machine Learning

Cognitive and mental representation is a cornerstone concept in understanding how minds—both human and artificial—process and interact with the world. In philosophy, it addresses how mental states correspond to external realities; in cognitive science, it explores how the brain encodes information, such as through spatial maps; and in machine learning, it refers to how models like large language models (LLMs) transform data into usable forms. This article examines the propensity of philosophers and cognitive scientists to propose theories of representation, highlights contrasting perspectives, and explores the relevance of these ideas to machine learning systems, particularly LLMs. It also discusses challenges in determining whether LLMs use statistical representations akin to biological cognitive maps and addresses problems with conceptual analysis of representation, drawing on David Papineau’s Is Representation Rife? (Papineau, 2004) and Stephen Stich’s What is a Theory of Mental Representation? (Stich, 1992), alongside insights from Judea Pearl and others.

Theories of Mental Representation in Philosophy and Cognitive Science

Philosophical Perspectives

In philosophy, mental representation is a central issue in the philosophy of mind, focusing on how mental states—such as beliefs, desires, and perceptions—relate to the world. The Representational Theory of Mind (RTM), rooted in ideas dating back to Aristotle, posits that mental states are semantically evaluable objects, like thoughts or concepts, that represent external states of affairs (Mental Representation, 2000). This view assumes that cognition relies on internal representations that carry meaning, enabling individuals to reason and act.

David Papineau, in his 2004 article Is Representation Rife?, argues that representations are ubiquitous in cognitive science and should be understood as biological phenomena analyzed in teleological terms—that is, in terms of their evolutionary functions (Papineau, 2004). He suggests that representations, such as a belief that “Santiago is east of Sacramento,” have content derived from their role in guiding behavior within biological systems. Papineau’s teleosemantic approach posits that representations are grounded in the brain’s operations, shaped by natural selection to serve specific purposes, like navigation or decision-making.

In contrast, Stephen Stich offers a skeptical view in his 1992 article What is a Theory of Mental Representation? (Stich, 1992). Stich argues that cognitive psychology does not need to rely on mental representations with semantic properties to explain cognition. He proposes a syntactic theory of mind, where mental states are characterized by their structural properties rather than their meaning or reference. Stich contends that semantic properties are extrinsic and problematic for causal-scientific explanations, suggesting that cognition can be understood through computational processes without invoking representations. This view aligns with eliminative materialism, which seeks to replace talk of mental states with descriptions of physical brain processes (Stich, 1983).

These contrasting perspectives highlight a broader debate in philosophy. Representationalists, like those supporting conceptual role semantics, argue that mental representations are defined by their functional roles in cognition (Field, 1978). Critics, including Stich and others like Daniel Dennett, question whether representations are necessary, proposing that cognition might be explained through dynamic or connectionist models that avoid semantic content (Dennett, 1987).

Cognitive Science and Neural Mechanisms

Cognitive scientists approach mental representation by studying how the brain encodes and processes information. A landmark discovery in this field was the identification of place cells in the hippocampus, neurons that fire when an animal occupies a specific location in its environment (O’Keefe & Dostrovsky, 1971). These cells are thought to form the basis of cognitive maps, mental representations of spatial layouts that enable navigation and memory.

Edward Tolman’s 1948 work, Cognitive Maps in Rats and Men, was pivotal in establishing this concept (Tolman, 1948). Through experiments with rats in mazes, Tolman demonstrated that rats could learn and navigate complex environments not through simple stimulus-response associations but by forming internal spatial maps. For example, rats trained in a maze could take shortcuts or detour around obstacles, suggesting they had a mental representation of the maze’s layout. Tolman’s findings challenged behaviorist theories, which emphasized observable behaviors over internal mental processes, and introduced the idea that cognitive maps are a fundamental aspect of spatial cognition.

Further research has identified grid cells in the entorhinal cortex, which fire in regular spatial patterns and likely track an organism’s movement through space (Moser et al., 2008). Together, place and grid cells form a neural system for spatial representation, supporting the idea that the brain constructs map-like structures to organize environmental knowledge. These findings have inspired parallels in machine learning, where researchers explore whether artificial systems can replicate such representational capabilities.

Contrasting Opinions on Representation

The debate over mental representation is marked by significant disagreement. Representationalists argue that mental states must be understood as representations with semantic content, essential for explaining how the mind processes information about the world. For example, Jerry Fodor’s psychosemantics posits that mental representations have meaning derived from their causal relationships with the environment (Fodor, 1987). Similarly, Ruth Millikan’s teleosemantic theory suggests that representations have content based on their evolutionary functions (Millikan, 1984).

In contrast, critics like Stich argue that positing mental representations may not add explanatory value. In his 1992 article, Stich suggests that cognitive science can rely on syntactic or connectionist models, where mental processes are explained by computational structures rather than semantic content (Stich, 1992). This view is supported by connectionist approaches, which model cognition as distributed patterns of neural activity rather than discrete representations (Ramsey et al., 1990). These contrasting views reflect a tension between those who see representations as central to cognition and those who advocate for alternative, non-representational frameworks.

Papineau’s position offers a middle ground, acknowledging the prevalence of representations but grounding them in biological and functional terms rather than abstract semantics. His teleosemantic approach aligns with cognitive science’s focus on neural mechanisms, bridging philosophical and empirical perspectives (Papineau, 2004).

Relevance to Machine Learning Systems

In machine learning, representation refers to how data is encoded and transformed within a model’s architecture, particularly in LLMs. These models, such as those powering chatbots or translation systems, rely on high-dimensional vector representations (embeddings) to capture semantic and syntactic relationships in language. These representations are learned statistically from large datasets, enabling tasks like text generation and comprehension (Bengio et al., 2013).

However, Judea Pearl has critiqued current machine learning systems for their limitations in reasoning about causality. In his 2018 paper Theoretical Impediments to Machine Learning With Seven Sparks from the Causal Revolution, Pearl argues that these systems operate in a statistical, model-free mode, which restricts their ability to reason about interventions and retrospection—key aspects of human intelligence (Pearl, 2018). He proposes that incorporating causal models, similar to those used in causal inference, could enable machines to achieve human-level intelligence by representing not just correlations but causal relationships.

Other researchers, like Yoshua Bengio, have also highlighted limitations in current representational learning. While LLMs excel at pattern recognition, their representations may not capture the deep, causal understanding required for general intelligence (Bengio et al., 2013). This gap suggests that insights from cognitive science, particularly about how the brain represents causal and spatial relationships, could inform the development of more advanced AI systems.

Challenges in Identifying Representation in LLMs

A significant challenge is determining whether the statistical representations in LLMs’ mid-tier nodes are analogous to biological cognitive maps or place cell activity. In the brain, place cells and grid cells form spatial representations that guide navigation, as seen in Tolman’s rat experiments and subsequent neuroscience research (Tolman, 1948; Moser et al., 2008). These representations are tied to specific neural mechanisms and physical locations, providing a clear framework for spatial cognition.

In contrast, LLMs’ representations are abstract, high-dimensional encodings of linguistic and conceptual relationships, not spatial environments. While these representations enable sophisticated language processing, they lack the spatial or causal grounding of biological cognitive maps. Determining whether LLMs’ mid-tier nodes form “representations” in the same sense as place cells is difficult because the two systems operate on different principles: biological representations are rooted in neural activity, while machine learning representations are emergent properties of statistical learning (Banino et al., 2018).

Recent research has explored whether neural networks can replicate cognitive map-like structures. For example, studies have shown that deep neural networks trained on spatial navigation tasks can develop representations resembling place and grid cell activity, suggesting some parallels between biological and artificial systems (Banino et al., 2018). However, applying this analogy to LLMs is less straightforward, as their representations are not spatial but linguistic, raising questions about whether they can be considered cognitive maps in any meaningful sense.

Cognitive Maps and Place Neurons as Analogies

Cognitive maps and place neurons provide a compelling analogy for understanding representation in both biological and artificial systems. In mammals, the hippocampal-entorhinal complex supports spatial navigation through place cells, which encode specific locations, and grid cells, which track movement patterns (Moser et al., 2008). Tolman’s experiments demonstrated that rats use cognitive maps to navigate mazes efficiently, suggesting a mental model of their environment (Tolman, 1948).

In machine learning, researchers have developed neural networks that mimic these spatial representations. For instance, Banino et al. (2018) showed that deep neural networks can learn grid-like representations for spatial navigation, resembling the activity of grid cells in the brain (Banino et al., 2018). Other studies have explored how neural networks can form cognitive maps of non-spatial domains, such as semantic spaces, by learning relationships between concepts (Wiskott & Munk, 2023). These findings suggest that artificial systems can replicate some aspects of biological representation, but LLMs’ focus on linguistic data limits direct comparisons to spatial cognitive maps.

Problems with Conceptual Analysis of Representation

Conceptual analysis of representation is fraught with challenges due to the term’s varied meanings across disciplines. In philosophy, representation often refers to the relationship between mental states and the world, emphasizing semantic content (Field, 1978). In cognitive science, it refers to specific neural mechanisms, like place cells or cognitive maps (Moser et al., 2008). In machine learning, representation is a mathematical concept involving data transformations, such as embeddings in LLMs (Bengio et al., 2013). These differing definitions create ambiguity, making it difficult to compare representations across fields without careful clarification.

This ambiguity complicates interdisciplinary discussions. For example, a philosopher might debate whether mental representations are necessary for cognition, while a machine learning researcher might focus on optimizing representational learning without considering philosophical implications. Bridging these perspectives requires a unified framework that acknowledges the diverse roles of representation in explaining cognition and building intelligent systems.

Conclusion

Cognitive and mental representation is a multifaceted concept that spans philosophy, cognitive science, and machine learning. Philosophers like Papineau and Stich offer contrasting views, with Papineau emphasizing the biological grounding of representations and Stich questioning their necessity. Cognitive scientists have identified neural mechanisms, like place cells and cognitive maps, that underpin spatial cognition, as demonstrated by Tolman’s pioneering work. In machine learning, LLMs rely on statistical representations to process language, but their limitations in causal reasoning, as noted by Pearl, highlight the need for more human-like models. Comparing LLMs’ representations to biological cognitive maps is challenging due to their abstract nature, and conceptual analysis of representation is complicated by disciplinary differences. As research progresses, interdisciplinary dialogue will be crucial for advancing our understanding of representation and its role in both natural and artificial intelligence.