Simons Foundation Launches Neural Computation Collaboration

Simons Foundation Launches Neural Computation Collaboration

In August 2025, the Simons Foundation unveiled the Simons Collaboration on the Physics of Learning and Neural Computation, led by Stanford’s Surya Ganguli. This major academic initiative combines physics, mathematics, theoretical neuroscience, and computer science to probe how large neural networks learn — with explicit attention to the differences between biological and artificial neural computation. The Collaboration directly addresses the question of whether biological neural dynamics possess computational properties that conventional artificial networks lack, and whether neuromorphic systems can capture these properties. This fundamental research has implications for consciousness theories, AGI development, and BCI technology. The initiative represents significant academic investment in understanding the computational principles that underlie both biological and artificial intelligence. See our neural networks vertical and consciousness research.

Research Program and Scope

The Simons Collaboration on the Physics of Learning and Neural Computation represents one of the most significant academic investments in understanding the computational principles that underlie both biological and artificial intelligence. The Collaboration brings together researchers from physics, mathematics, theoretical neuroscience, and computer science to probe fundamental questions: How do large neural networks learn? What computational principles explain the brain’s remarkable efficiency? Do biological neural dynamics possess properties that current artificial networks cannot replicate? And can neuromorphic systems capture these properties?

Led by Stanford’s Surya Ganguli — a physicist who has become one of the most influential theoretical neuroscientists studying the intersection of statistical mechanics, neural network theory, and biological computation — the Collaboration leverages the tools of theoretical physics (statistical mechanics, information theory, dynamical systems theory, random matrix theory) to understand how learning emerges in complex neural systems.

The Collaboration involves researchers from multiple institutions including Stanford University, Columbia University, the Institute for Advanced Study, and leading European neuroscience centers. Participants include both theorists developing mathematical frameworks and experimentalists providing empirical data on biological neural computation. This interdisciplinary approach is essential because the questions being addressed — how learning works, what makes biological computation efficient, whether artificial systems can replicate biological capabilities — span disciplinary boundaries.

Key Research Questions

Learning Theory: How do neural networks — both biological and artificial — learn from data? Standard deep learning theory focuses on optimization (how gradient descent finds good solutions) and generalization (why networks trained on finite data perform well on unseen data). The Collaboration extends these questions to biological learning, where the optimization process (synaptic plasticity, not backpropagation) and the generalization mechanism (ecological adaptation, not i.i.d. sampling) differ fundamentally from artificial learning. Understanding these differences could inspire new training algorithms for AI that are more efficient, more robust, and more biologically plausible.

Biological Computation Efficiency: The human brain operates on approximately 20 watts while performing cognitive feats that require megawatts in artificial data centers. What computational principles account for this five-to-six orders of magnitude efficiency gap? The Collaboration investigates whether biological efficiency arises from spiking communication, sparse coding, local learning rules, metabolic constraints, architectural organization, or some combination of these factors. Understanding the source of biological efficiency could inform the design of more efficient AI hardware and algorithms.

Representation Learning: How do neural networks develop internal representations of the external world? In artificial neural networks, representations emerge through gradient-based optimization of a training objective. In biological neural networks, representations emerge through interaction with the environment, guided by evolutionary priors, developmental programs, and synaptic plasticity rules. Comparing the representations learned by biological and artificial systems could reveal whether current AI approaches capture all relevant aspects of cognition or miss fundamental representational capabilities.

Implications for AI Architecture

The Collaboration’s research has direct implications for neural network architecture design:

Beyond Backpropagation: If biological learning rules (spike-timing-dependent plasticity, reward-modulated plasticity, dendritic computation) provide computational advantages that backpropagation lacks, this would motivate the development of new training algorithms that incorporate biological insights. Several Collaboration researchers are developing “biologically plausible” learning algorithms that maintain gradient-based optimization’s effectiveness while respecting biological constraints like local learning, forward-only computation, and spike-based communication.

Architecture Search Through Physics: The physics-based approach to understanding neural computation could provide principled methods for architecture search — determining the optimal neural network architecture for a given task not through expensive trial-and-error but through theoretical analysis of the task’s computational requirements.

Scaling Laws and Phase Transitions: The Collaboration applies statistical mechanics tools to understand scaling laws and phase transitions in neural network learning. Just as physical systems undergo sudden phase transitions (water to ice, paramagnet to ferromagnet), neural networks may undergo computational phase transitions as they scale — potentially explaining the emergent capabilities observed in frontier AI systems.

Connection to Consciousness Research

The Collaboration’s research has implicit connections to consciousness research. If biological neural dynamics possess computational properties that artificial networks lack — properties related to temporal coding, oscillatory dynamics, synaptic integration, or dendritic computation — these properties could be precisely the ones that give rise to conscious experience. Understanding the computational differences between biological and artificial neural systems could reveal which features are computationally necessary (and potentially sufficient) for consciousness.

Integrated Information Theory predicts that biological neural networks have higher consciousness (higher Phi) than artificial neural networks due to their denser recurrent connections and richer causal structure. The Collaboration’s mathematical tools could provide more rigorous analysis of this prediction than currently available methods.

Global Workspace Theory predicts that consciousness requires specific information processing dynamics (ignition, broadcasting, capacity limitation) that may depend on neural timing precision and oscillatory coupling. If the Collaboration identifies these dynamics as computationally essential features of biological networks, this would constrain which artificial architectures could support consciousness.

Institutional Significance

The Simons Foundation — one of the world’s largest private funders of basic research, founded by mathematician and hedge fund pioneer James Simons — has a track record of investing in fundamental research programs that produce transformative scientific insights. The Foundation’s decision to invest in neural computation research signals that the scientific community views the computational principles of intelligence as one of the most important frontiers in science.

The Collaboration complements other major research initiatives in the field, including the European Human Brain Project (which developed the BrainScaleS neuromorphic platform), the BRAIN Initiative (which funded neural recording technology advances), and the Wellcome Trust’s consciousness research program. Together, these initiatives are creating a comprehensive research infrastructure for understanding biological and artificial intelligence.

See our neural networks vertical and consciousness research.

Research Agenda and Expected Outputs

The Collaboration’s research program spans several interconnected areas with direct implications for AI development and consciousness research. First, understanding why biological neural networks generalize from limited training examples — while artificial networks require orders of magnitude more data — could reveal architectural principles that improve the sample efficiency of deep learning systems. Second, investigating the role of temporal dynamics (oscillations, spike timing, synaptic plasticity) in biological learning could inform the design of neuromorphic computing architectures that capture computational properties absent from current artificial systems. Third, analyzing the mathematical structure of learning in biological versus artificial networks could determine whether fundamental differences exist that cannot be bridged through scaling alone — a question directly relevant to the AGI timeline debate.

Expected research outputs include publications in theoretical neuroscience and machine learning, new mathematical tools for analyzing learning dynamics, open-source software for neural circuit modeling, and potentially new architectural principles for neural network design. The Collaboration’s 5-year funding commitment provides the sustained support needed for fundamental research that may not produce immediate applications but could reshape the theoretical foundations of both neuroscience and AI.

Implications for the AI Industry

For the $390.9 billion AI market, the Simons Collaboration represents a strategic investment in understanding whether biological neural computation possesses properties that current artificial systems lack. If the Collaboration discovers that biological learning relies on computational mechanisms not captured by backpropagation-trained transformers — such as dendritic computation, spike-timing-dependent plasticity, or oscillatory binding — this would strengthen the case for neuromorphic computing approaches and potentially redirect billions of dollars in AI research investment.

For brain-computer interface development, the Collaboration’s research on neural computation principles could improve neural decoding algorithms by providing deeper understanding of how neural populations encode information. Decoders based on accurate computational models of neural dynamics would outperform purely data-driven approaches, particularly for complex applications like speech restoration where understanding the neural control of articulation is essential for accurate decoding.

For consciousness research, the Collaboration’s investigation of the differences between biological and artificial neural processing directly informs the question of whether artificial systems could possess consciousness-relevant properties. If consciousness depends on computational mechanisms specific to biological neural circuits — as some researchers argue — then the Collaboration’s findings would constrain the space of possible artificial consciousness substrates.

See our neural networks vertical and consciousness research.

The Physics of Learning: A New Paradigm

The Collaboration’s emphasis on the “physics of learning” reflects a growing recognition that the mathematical tools of theoretical physics — statistical mechanics, dynamical systems theory, information theory, random matrix theory — are uniquely suited to understanding how neural networks learn. Just as statistical mechanics explains how macroscopic properties of matter emerge from microscopic interactions between atoms, the physics-of-learning framework seeks to explain how macroscopic cognitive capabilities emerge from microscopic interactions between neurons or artificial processing units. This approach has already yielded insights into the geometry of loss landscapes (the mathematical surfaces that learning algorithms navigate during training), the role of symmetry and symmetry-breaking in neural representation learning, and the statistical mechanics of generalization (why networks trained on finite data can generalize to unseen examples). For the $34.28 billion deep learning market, these theoretical advances have direct practical implications — understanding why networks generalize enables the design of architectures and training procedures that generalize more efficiently, reducing the data and compute requirements that currently constrain AI development.

Cross-Disciplinary Research Infrastructure

The Collaboration brings together researchers from disciplines that rarely interact directly — theoretical physics, pure mathematics, experimental neuroscience, and machine learning engineering. This cross-disciplinary infrastructure is essential because the fundamental questions about neural computation cannot be answered within any single discipline. Understanding why biological neural networks learn efficiently from limited data requires the mathematical tools of theoretical physics, the empirical methods of experimental neuroscience, the algorithmic knowledge of machine learning, and the analytical frameworks of pure mathematics. The Collaboration’s institutional structure — with regular workshops, visiting researcher programs, and collaborative research projects — creates the sustained interaction needed for genuine cross-disciplinary breakthroughs rather than superficial interdisciplinary gestures. For the $390.9 billion AI market and the $2.94 billion BCI market, the Collaboration represents the kind of fundamental research investment that produces transformative insights with timescales measured in years rather than quarters — insights that commercial R&D cannot generate but that commercial applications depend upon.

The Broader Context of Fundamental AI Research

The Simons Collaboration exists within a broader landscape of fundamental research on intelligence that includes the European Human Brain Project, the US BRAIN Initiative, the Allen Institute for Brain Science, and numerous university research programs. What distinguishes the Simons Collaboration is its explicit focus on the mathematical principles underlying learning — not the engineering of specific AI systems or the mapping of specific brain circuits, but the abstract mathematical laws that govern how any information processing system learns from experience. If these laws exist and can be discovered, they would provide a unified theoretical framework for understanding biological intelligence, artificial intelligence, and the relationship between them — with implications for every market segment from cognitive computing to consciousness research to brain-computer interfaces. The Collaboration’s five-year funding commitment from the Simons Foundation provides the sustained support that this ambitious research program requires.

The Significance for Artificial Intelligence Safety

The Simons Collaboration’s research has direct implications for AI safety and alignment. Understanding the fundamental mathematical principles governing learning in neural networks could reveal whether AI systems are converging toward human-like cognitive architectures as they scale, or whether they are developing fundamentally alien forms of intelligence. If the physics of learning produces universal principles that biological and artificial systems must both obey, then studying biological learning provides insights into the future behavior of AI systems that purely empirical AI research cannot. Conversely, if biological and artificial learning follow different mathematical laws, this would suggest that analogies between AI and human cognition are misleading and that safety research must develop AI-specific analytical frameworks rather than borrowing from neuroscience. For the AGI governance community, these fundamental questions about the nature of learning have direct policy implications, determining whether biological intelligence serves as a reliable model for predicting and controlling artificial intelligence or whether new conceptual frameworks are needed.

Updated March 2026. Contact info@subconsciousmind.ai for corrections.