- The paper demonstrates that predictive success in AI does not guarantee a mechanistic understanding of neural processes.
- It details how foundation models applied to neural data achieve high accuracy yet often overlook underlying biological mechanisms.
- It proposes integrating biological constraints and mechanistic interpretability to bridge the gap between prediction and explanation.
Foundation Models in Brain Science: From Predictive Power to Mechanistic Understanding
Introduction
This paper critically examines the integration of AI foundation models—large-scale, pretrained neural networks—into brain science, with a focus on their capacity to move beyond predictive accuracy toward mechanistic explanation. The authors analyze recent advances in neural and behavioral foundation models, such as those trained on calcium imaging and human decision-making data, and interrogate the extent to which these models can illuminate the computational principles underlying neural activity and cognition. The central thesis is that predictive success, while necessary, is insufficient for scientific understanding; mechanistic interpretability and theoretical grounding are essential for foundation models to contribute meaningfully to neuroscience.
Foundation Model Paradigm and Its Extension to Neuroscience
The foundation model paradigm is characterized by large-scale self-supervised pretraining, typically using transformer architectures, followed by domain-specific finetuning. This approach has yielded state-of-the-art results in NLP, vision, speech, and multimodal domains, and is now being applied to neural and behavioral data. The shared tokenization principle enables the extension of SSL objectives across modalities, allowing models to learn versatile representations from raw data. Empirical scaling laws demonstrate that increasing data, compute, and model parameters reliably improves predictive performance, motivating the deployment of ever-larger models in scientific domains.
In neuroscience, foundation models have demonstrated strong predictive accuracy in tasks such as neural response prediction, behavioral modeling, and clinical variable estimation. For example, a neural foundation model trained on mouse visual cortex calcium imaging data generalizes across stimulus domains and individual animals, capturing information about cell types and connectivity. Similarly, the Centaur behavioral model predicts human choices across diverse experimental paradigms, outperforming classical cognitive models on held-out participants and novel tasks.
Prediction Versus Explanation: Mechanistic Gaps
Despite their predictive prowess, foundation models often fall short of providing mechanistic insight. The distinction between prediction and explanation is emphasized through several case studies. Models may exploit statistical regularities or shortcuts rather than uncovering the true generative processes underlying neural or cognitive phenomena. For instance, models trained on synthetic physics data can achieve high accuracy without internalizing the governing physical laws, and behavioral models may align with human data superficially while diverging in well-established psychological experiments.
The authors highlight that even when model representations partially correspond to biological features (e.g., neuron types, dendritic morphology), this does not constitute a mechanistic account of computation. The risk is that foundation models may simply replace one black box (the brain) with another (a deep neural network), without advancing explanatory understanding.
Mechanistic Interpretability and Theoretical Grounding
Recent advances in mechanistic interpretability offer pathways to bridge the gap between prediction and explanation. Analyses of attention weights, hidden-layer activations, and functional subcircuits in transformer models reveal compositional structures that can be mapped to specific computations. These algorithmic circuits, while distinct from biological ones, echo canonical microcircuit motifs in neuroscience. At the representational level, both specialized units (e.g., "cat neurons," "concept cells") and distributed semantic patterns have been identified, paralleling debates in neural coding.
Intervention techniques, such as activation engineering and vector steering, demonstrate that semantic relationships in LLMs are encoded as consistent geometric patterns in high-dimensional spaces. This leads to testable hypotheses about the organization of human conceptual knowledge, suggesting that similar linear operations may be recoverable in neural activity patterns.
However, the authors caution that scaling laws in AI do not guarantee convergence on biological mechanisms, due to the absence of evolutionary and developmental constraints that shape biological systems. Embedding such constraints into model architectures and optimization procedures may enhance the explanatory value of foundation models.
Limitations, Opportunities, and Future Directions
The paper identifies several limitations in the current application of foundation models to brain science:
- Predictive alignment does not imply mechanistic fidelity; models may fit data without capturing underlying processes.
- Shortcut learning and reliance on spurious correlations undermine explanatory claims.
- Lack of evolutionary constraints in model development limits biological plausibility.
Opportunities for progress include:
- Mechanistic interpretability research to map model computations to functional subcircuits and generate experimentally testable hypotheses.
- Integration of biological constraints into model architectures and training regimes.
- Digital twins for in silico experimentation, personalized medicine, and neurotechnology, contingent on achieving mechanistic realism.
The authors argue that foundation models can serve as theory-bearing instruments if their computations are grounded in neuroscience and psychological theory, and if their internal mechanisms are made transparent and experimentally testable.
Conclusion
Foundation models have demonstrated substantial predictive power in neuroscience and cognitive science, but their scientific value depends on moving beyond data fitting to mechanistic explanation. Mechanistic interpretability, theoretical grounding, and experimental validation are essential for these models to advance understanding of brain and cognition. The transformation of foundation models from predictive tools to explanatory instruments will determine their impact on the future of neuroscience.