"More Is Different'' in Neural Circuits: Algebraic Emergence of Effective Theories in Canonical Recurrent Motifs of Biological Neuronal Networks
This paper demonstrates that composing canonical neural circuit motifs, such as winner-take-all and divisive normalization, can algebraically generate emergent effective theories and reversible group dynamics within their transition monoids that are absent in the individual aperiodic components, thereby revealing that the computational repertoire of recurrent circuits is fundamentally constrained and defined by the algebraic structure of their compositional interfaces.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The brain is a machine built from repeating patterns. Just as a house is constructed from bricks, beams, and windows, the neural circuits inside our skulls are assembled from a small set of standard building blocks. Two of the most common blocks are known as divisive normalization and winner-take-all competition. The first acts like a volume knob that turns down the overall loudness of a group of neurons when they are all firing too loudly, ensuring that no single signal drowns out the others. The second acts like a referee in a crowded room, allowing only one group of neurons to speak while silencing the rest. For decades, neuroscientists have understood these blocks as functional tools: one rescales activity, the other selects a winner. But a new study asks a deeper question about what happens when these blocks are connected. It investigates whether the connection itself creates something entirely new, a capability that neither block possesses on its own.
This research, led by Nima Dehghani at the Massachusetts Institute of Technology, treats these neural circuits not just as biological systems, but as mathematical machines. The goal was to see if the way these machines are wired together could generate a form of computation that is impossible to find in any single part. To do this, the researchers simplified the complex, continuous firing of real neurons into a set of discrete states, like a light switch that is either on or off. They then built digital models of the normalization and competition circuits and ran them through every possible sequence of inputs. Instead of watching a single path of activity, they mapped the entire landscape of what the circuits could do when the inputs changed over time. They were looking for a specific kind of behavior: a cycle where the system could return to a previous state, creating a loop of reversible action. In the language of the study, they were searching for a "group" structure, a mathematical signature of a system that can remember and cycle through states, rather than just dissolving into a fixed point.
The study began by testing the two circuits in isolation. When the researchers looked at the winner-take-all circuit alone, they found that under any single, unchanging input, the system was purely dissipative. It would always settle down into a single, stable state where one group of neurons won and the rest were silent. There were no loops, no cycles, and no way to return to a previous configuration once the system had moved forward. The same was true for the normalization circuit in its stricter, more complex form. Every single rule that governed these circuits, when applied alone, drove the system toward a dead end. However, when the researchers combined these circuits, the story changed. They discovered that by switching between different inputs in a specific sequence, they could force the system into a loop. Even though every individual step was a one-way street that erased memory, the sequence of steps created a hidden path that allowed the system to return to where it started. This was the first surprise: a system made entirely of irreversible parts could, through composition, generate a reversible loop.
The most significant finding emerged when the researchers connected the two circuits in a specific direction. They created a setup where the winner of the competition determined the settings for the normalization process. In this arrangement, the selected winner did not just sit passively; it actively reshaped the environment for the next moment. When they analyzed the combined system, they found a cycle that belonged to neither circuit alone. In this loop, the identity of the winning group of neurons remained constant, but two other things changed together: the state of the inhibitory gate that controlled the winner, and the normalization setting that governed the overall gain. The system would toggle between a state where the winner was active and the gate was open, and a state where the winner was active but the gate was closed, all while the normalization setting shifted in perfect sync. This cycle was a genuine composite structure. It was not a feature of the competition block, nor was it a feature of the normalization block. It was a new property that existed only because the two blocks were coupled in this specific way.
To ensure this result was not a fluke of their specific model, the researchers tested thousands of variations. They tried connecting the circuits in the opposite order, where normalization set the stage for competition, and they found that while complex structures still existed, the most direct and obvious cycle was buried deep within the system, hidden behind simpler, local loops. They also tested every possible way the two circuits could be wired together, mapping out 65,536 different interface configurations. They found that the composite cycle appeared in nearly all of them, provided the connection allowed the winner to influence the normalization settings. They also tested whether the timing of the updates mattered. When the circuits were updated simultaneously, the cycle appeared as a hard-coded feature of the update rule itself. But when the updates happened in sequence, the cycle emerged purely from the interaction, proving that the structure was a result of the composition, not a pre-existing artifact.
The study concludes that the algebra of neural computation is richer than the sum of its parts. It demonstrates that the brain's ability to compute is not just about the individual operations of its circuits, but about how those operations are chained together. By choosing how to connect a normalization block to a competition block, the brain can effectively program a new degree of freedom. This new freedom allows the system to maintain a stable winner while simultaneously modulating its own sensitivity, creating a dynamic state that is more flexible than either block could achieve alone. The research suggests that the "more is different" principle, often applied to the complexity of the whole brain, also applies to the smallest building blocks of neural circuits. When these blocks are composed, they generate a repertoire of behaviors that cannot be predicted by looking at the blocks in isolation. The transition from a simple, dissipative system to one capable of reversible, cyclic computation is not a mystery of biology, but a mathematical certainty of composition. The brain, in this view, is not just a collection of parts, but a programmable system where the wiring itself writes the code.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.