Brain Criticality - Optimizing Neural Computations

The core claim behind “brain criticality” is that neural networks operate near a second-order phase transition—an edge-of-instability regime where activity becomes scale-free and correlations spread unusually far. That matters because the same critical tuning is predicted to maximize how effectively the brain transmits information, balancing responsiveness to inputs against runaway excitation.

The explanation starts with phase transitions in physics. In everyday first-order transitions like boiling, a system jumps between phases with a discontinuous change in an order parameter. Second-order transitions behave differently: the order parameter changes continuously, yet the system develops a special “critical point” where intermediate behavior emerges. At that point, latent heat disappears (in the water example, the liquid–gas boundary becomes a supercritical fluid), and the system shows properties that don’t depend on a single characteristic scale.

To make the idea concrete, the transcript uses the Ising model, a lattice of spins that can be +1 or −1. Neighboring spins prefer to align (captured by an interaction energy proportional to the negative product of spins), while temperature adds stochastic fluctuations via the Boltzmann distribution. At low temperature, alignment dominates and magnetization is high; at high temperature, randomness dominates and magnetization collapses. Near the critical temperature, neither tendency wins outright. The result is a peak in dynamic correlation: spins fluctuate in coordinated ways, and the correlation length grows sharply, meaning distant parts of the system move together more than they do away from criticality.

Criticality also produces scale-free structure. Snapshots of the Ising lattice look statistically similar across zoom levels, and cluster-size statistics follow a power law rather than an exponential. In the transcript’s example, the probability of observing a cluster of size x versus 2x changes by a constant factor independent of x, which is the hallmark of scale invariance. In log-log coordinates, power laws appear as straight lines, and many critical observables share this kind of scaling.

The neuroscience bridge comes from experiments on “neuronal avalanches.” In 2003, John Beggs and Dietmar Plenz reported that spontaneous activity in rat somatosensory cortex neuron cultures—recorded with an 8 by 8 electrode grid—appears in cascades separated by quiescent periods. When avalanches are defined by duration and size (number of electrodes crossing a threshold, sometimes including amplitude), their distributions follow power laws, implying no characteristic scale and suggesting proximity to a second-order phase transition. Similar power-law avalanche behavior has since been reported across species and recording modalities, from single neurons to EEG.

To map physics concepts onto neural dynamics, the transcript reframes the system as a branching process: each active neuron probabilistically activates downstream neurons, plus a small chance of spontaneous activation. A single control parameter, the branching ratio σ (the sum of outgoing transmission probabilities, equal to the average number of descendants per active ancestor), determines the regime. For σ < 1, activity dies out; for σ > 1, activity amplifies and can resemble epileptiform runaway; at σ = 1, activity neither decays nor explodes on average, yet avalanche sizes and durations remain power-law distributed. In real brains, σ is shaped by the balance of excitation and inhibition, and pharmacologically blocking inhibition disrupts the normal power-law pattern toward supercritical behavior, while blocking excitation pushes toward subcritical dynamics.

Finally, the transcript argues that critical tuning improves information transmission. In a simplified “guessing game,” outputs are uninformative when activity dies out (subcritical) and uninformative when activity saturates (supercritical). At σ = 1, output activity most reliably reflects input activity, producing a peak in an information-transmission measure analogous to the correlation peak at the Ising critical temperature. The overall picture is that hovering near criticality may be an evolved strategy to maximize computational capability while avoiding both silence and runaway excitation.