Adding inhibition and homeostasis cut variability 13-fold in a three-population model of V1

Chaos has a narrow technical meaning here, and Borjkhani drew the line himself. "In this context, chaos does not mean ordinary noise or disorder. It is a deterministic form of dynamics in which a very small difference at the beginning can rapidly take the entire system in a different direction. For the brain, this is a potential source of flexibility, but also a risk to stable information processing," he said ^[6]. The paradox the paper sets up follows from that definition: a rigid visual system could not adapt to unexpected input, and an uncontrolled chaotic one would turn identical sensory input into entirely different outputs ^[10].

The reduction from 0.325 to 0.024 works out to 92.6 percent ^[1], which the publisher's summary rounds and presents as up to 93 percent ^[7]. Put the other way round, the stabilised network keeps about one thirteenth of the variability it had before the loops were added ^[2].

The design choice that makes this legible is the size of the state space. Three variables stand in for the whole of V1: pyramidal cells, which the authors put at roughly 80 percent of cortical neurons ^[8], fast-spiking PV+ interneurons, and a slower modulatory term folding together thalamic and neuromodulatory input. It all runs on a modified Lotka-Volterra scaffold ^[2]. Chaos was the model's usual condition. Nearly 90 percent of 225 parameter combinations produced it, around 200 of the 225 ^[3][3]. In a three-variable system you can add one feedback loop, take it away, and attribute the change in variability to that loop.

The check that the stabilised circuit can still see is the orientation selectivity index. The virtual neurons scored 0.31 to 0.38, inside macaque benchmarks, and separated preferred edge orientations slightly more clearly under controlled near-threshold dynamics ^[5]. OSI is a summary of how sharply a cell's firing peaks at one edge angle. Matching it shows the stabilised network did not go blind to orientation. It does not show that the model's moment-to-moment behaviour resembles a real V1 neuron's. The paper's positive claim on that front is the weaker one, that moderately irregular chaotic dynamics did not degrade visual processing ^[9].

So this is a sufficiency argument. Two feedbacks with known biological counterparts, fast interneuron inhibition and slow homeostatic drive, are enough inside this model to hold a chaotic three-population system near the boundary where it stays responsive ^[4]. Whether the cortex actually needs either loop is a separate question, and this model cannot settle it. The summary also reports 0.325 and 0.024 without stating what the variability measure is, so the 93 percent is a comparison in the model's own units ^[11].