- The paper introduces the Umwelt Representation Hypothesis as an alternative to the global optimality perspective in cognitive systems.
- It demonstrates that neural and artificial representations diverge due to overlapping ecological and functional constraints rather than converging to a unique global optimum.
- Empirical evidence from cross-species, individual, and ANN studies supports systematic alignment clusters, urging targeted comparisons in model design.
The Umwelt Representation Hypothesis: Reframing Representational Alignment in Brains and Artificial Neural Networks
Introduction
The prevailing narrative in computational neuroscience and the neuroconnectionist paradigm holds that artificial neural networks (ANNs) and biological brains, when exposed to sufficient data and constraints, tend to develop convergent internal representations of the world. These findings have inspired the hypothesis of representational Universality, positing that all sufficiently capable cognitive systems converge on the same optimal internal model of objective reality. "The Umwelt Representation Hypothesis: Rethinking Universality" (2604.17960) rigorously interrogates this claim, introducing the Umwelt Representation Hypothesis (URH) as a theoretically and empirically grounded alternative. The authors argue that representational alignment is not tantamount to convergence on a unique global optimum, but rather emerges from overlapping ecological and functional constraints—Umwelts—that systematically shape neural and artificial representations.
Universality and the Assumption of Global Optimality
The Universality hypothesis interprets representational alignment as indicative of convergence on shared, globally optimal representations dictated by the structure of the external world. Drawing from metaphysical realism, the Platonic Representation Hypothesis asserts that cognitive systems—regardless of architecture, training regimen, or species—are driven toward a common veridical world model, especially as their capacity and data exposure increase. This view is exemplified by the analogy that "all roads lead to Rome" and strengthened by findings showing similar representational geometries not only between diverse ANNs and humans but also across input modalities such as vision and language [13,16,22,23].
Under this perspective, misaligned or dissimilar representational components are treated as idiosyncratic errors or incomplete optimization (the "Anna Karenina scenario"): all systems that closely approximate the global optimum should be alike, and all deviations reflect noise or failure rather than adaptive specialization.
The Umwelt Representation Hypothesis
The URH, drawn from the ecological and phenomenological concept of Umwelt, presents an alternative explanatory framework. Each biological or artificial agent operates within a bounded Umwelt, structured by sensory apparatus, effector capabilities, goals, environmental pressures, and internal constraints (both inherited and acquired). Representational content is shaped by these constraints, resulting in systems that filter and compress information relevant to their survival or task performance.
The URH does not deny the existence of overlapping or even universally shared representations but reinterprets their origin: alignment emerges where Umwelts overlap, not because of movement toward a single optimal representational solution. Systematic differences—across species, individuals, or models—are often adaptive, reflecting local optima in an evolutionary or task-specific context.
Consequently, rather than expecting convergence on a unitary "world model," the URH predicts partial alignment: clusters of shared representational structure occur when and where ecological constraints overlap, and divergence otherwise. Even with increased model capacity and exposure, meaningful differences are expected to persist if the underlying constraint structure remains distinct.
Empirical Evidence for Systematic Representational Divergence
Experimental and observational studies reveal robust, structured representational differences at multiple scales:
- Across species: Neural systems reflect profound evolutionary specialization (e.g., echolocation in bats, diverse photoreceptors in mantis shrimp, or magnetoreception in birds) that cannot be explained by incomplete convergence toward an objective optimum. Connectivity patterns mirror evolutionary hierarchies, underscoring constraint-driven divergence.
- Within species and individuals: Systematic individual differences—cultural variants in visual perception, language-specific neural circuits, or acquired expertise (e.g., face, Pokémon, or car recognition)—are prevalent. These differences are structured, stable, and often beneficial rather than stochastic noise.
- Between humans and ANNs: Detailed behavioral and neuroscientific analyses reveal that standard ANNs, despite high levels of performance, often rely on representational strategies (e.g., texture bias vs. shape bias in visual recognition) distinct from those of humans. Improvements in model accuracy do not always correspond to improved brain alignment, and different training objectives, network depths, or initializations induce systematic differences among ANNs themselves.
This convergence/divergence landscape defies the Universality prediction that only optimal representations are shared and all others constitute unstructured noise. Instead, it substantiates the URH claim that systematic, meaningful differences originate in non-overlapping ecological constraints.
Implications for Model Comparison and the Study of Cognition
The emergence of substantial representational alignment between human brains and ANNs has prompted concerns about the diminishing discriminative power of model comparison for understanding brain function under the Universality framework. The URH provides a methodological counterpoint: representational comparisons should be interpreted as probing the structure of ecological constraint space, revealing clusters of alignment corresponding to shared Umwelts.
This perspective entails two major shifts:
- Mapping Alignment Clusters: Rather than searching for a single "most brain-like" model, research should focus on identifying and characterizing clusters of systems (brains or ANNs) that align under specific sets of constraints. This approach is inherently comparative and can reveal the hierarchical organization of cognition across species, modalities, and model classes.
- Systematic Exploration of Constraints: The authors advocate for systematic variation of model architecture (beyond CNNs and transformers to include, e.g., recurrent, topographic, or predictive coding models), training data (diversifying datasets and task regimes), objective functions, and cognitive task paradigms. Such experimentation will illuminate the boundaries and topology of alignment clusters, enabling the disambiguation of human-centric and species-specific factors from those that are universal or pan-modal.
The emphasis on ecological constraint space also demands improvements in representational alignment metrics. Current flexible parametric models may obscure systematic differences; sharper, more theory-driven metrics will be necessary to distinguish truly shared components from superficial or artifactually aligned representations.
Theoretical and Practical Implications
The adoption of URH reshapes foundational debates in computational neuroscience, cognitive science, and the development of artificial intelligence. Conceptually, it weakens the case for a metaphysically realist, one-world-model approach in favor of pragmatist or ecological models that prioritize utility, adaptivity, and action-relevance over representational correspondence. Practically, it casts doubt on the feasibility, necessity, or even desirability of constructing foundation models that seek to amortize all possible Umwelts; instead, targeted, ecologically valid models may offer a better path toward explainable and robust AI.
For neuroscience, URH motivates more nuanced cross-species and cross-individual comparisons, urging the field to embrace representational diversity as a source of insight into adaptive specialization rather than as mere noise. For AI, it frames the challenge of alignment not as an asymptotic convergence to human intelligence, but as the principled engineering of systems tailored to well-defined constraints and functions.
Conclusion
The Umwelt Representation Hypothesis (2604.17960) offers a compelling, constraint-based reframing of representational alignment across brains and artificial neural networks. By grounding representational content in ecological and physiological specificity, URH explains both convergence and divergence as adaptive, systematically structured responses to local pressures rather than as movement toward or away from a unitary global optimum. This approach preserves the scientific utility of model comparison, enriches our understanding of cognitive diversity, and has critical implications for the design and interpretation of future AI systems. The conceptual and methodological pivot advocated by the authors is poised to guide research toward a richer, more explanatory account of cognition—not only in machines, but across the biological spectrum.