This work sits at the fringe of the knowledge map: almost nothing nearby, with the nearest neighbor at distance 0.557. Nevertheless, surrounding works concern the brain, recurrent neural networks, and synapses: how the brain remembers after silence, how genes learn to play in time with chaos, why neurons cannot be friends with everyone. The common thread among these neighbors is an attempt to understand how connection structure gives rise to collective behavior. This paper fits into that thread but examines a parameter that neighbors did not touch: how precisely incoming signals to each neuron cancel each other. Previous works studied balance only as a way to remove outliers in the connectivity matrix spectrum, without asking whether it changes the dynamics itself.
The authors took the classical random recurrent network model of Sompolinsky and added one parameter: the degree of local balance. Each connection weight was redefined so that the sum of incoming weights to a node tends to zero under full balance. They then numerically simulated networks of one thousand and three thousand nodes with different activation functions: linear ReLU, quadratic superlinear, sigmoid, and sublinear. For analytical description, they used dynamical mean-field theory, deriving self-consistent equations via path integrals. Stability was assessed by the fraction of networks converging to a fixed point and by the maximal Lyapunov exponent.
The effect of balance turned out to be opposite for different classes of activation functions. For linear ReLU and quadratic functions, balance suppressed divergence: under full balance, the critical connection strength after which activity explodes increased from 1.4 to 1.8. For sigmoid and sublinear functions, balance instead induced chaos: the maximal Lyapunov exponent grew monotonically with increasing balance. For odd functions, such as hyperbolic tangent, balance had no effect on dynamics. The reason is that for odd functions, the population-averaged activity is zero, and the self-generated input that balance removes vanishes by itself. For saturating functions, another surprise emerged: the effective dimensionality of dynamics, measured by the participation ratio, changed non-monotonically with increasing balance, peaking at an intermediate value.
The work shows that local balance of connections is not just a cosmetic fix to the spectrum but an independent control parameter of collective dynamics. It can either stabilize the network or throw it into chaos, depending on how the individual element is structured. This challenges the common assumption of odd activation functions, which masked the effect. For neuroscience, the result means that the excitation-inhibition balance observed experimentally in the cortex may directly influence whether the network is in a quiet or chaotic activity regime.
The topic could develop toward validation on real data. Large-scale connectomic datasets combining synaptic maps with activity recordings would allow estimating local balance by the number and size of synapses and comparing it with the dimensionality of population activity. An open question remains how well anatomical balance corresponds to effective balance of forces in the model, since a synapse's influence depends on its location on the dendrite and cell type. It would also be interesting to extend the analysis to networks with sparse connections and more realistic single-neuron dynamics.
The results directly affect computational neuroscience, the theory of recurrent neural networks, and the physics of disordered systems. They may also be useful in ecology and genetics, where similar models describe species communities and gene networks.
The immediate next step is to test predictions on functional connectomics data, comparing anatomical balance with activity dimensionality. In theory, one should investigate how balance interacts with other structural features, such as sparsity or low-rank components.
The work connects to a long-standing problem in neuroscience: why cortical networks operate in a regime close to chaos but do not tip into it. The local excitation-inhibition balance studied here may be one mechanism keeping the network at the edge. It also contributes to understanding how anatomy determines computational properties of networks, which remains an open question.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors took the classical Sompolinsky model of a random recurrent network and added a local balance parameter. Numerically and analytically, they showed that balance suppresses divergence for linear and superlinear activation functions, but induces chaos for saturating and sublinear ones, and has no effect on odd functions. They found a nonmonotonic change in effective dimensionality at intermediate balance.
Where its strength is
The work introduces local balance as an independent control parameter of dynamics, not just as a regularization of the spectrum. This opens a simple way to switch network activity regimes by changing a single parameter.
Where it could go next
Worth trying
Why this may matter
Excitation-inhibition balance is an experimentally observed property of cortical networks. Understanding how it governs dynamics could link structural measurements to functional brain states and help design recurrent neural networks with desired activity regimes.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.54, against a usual 0.64. It looks like few have drilled nearby.
Under full balance (when the sum of incoming weights is zero), the time-averaged activity of each neuron in the network is strictly zero, even if the network is in a chaotic regime. This is analytically proven in the appendix of the paper.