Local homeostatic plasticity can hold deep networks near criticality

What the work claims

The authors claim that the global dynamical state of a deep neural network can be regulated by a purely local homeostatic plasticity rule that never measures any network-wide property.¹ Neurons with activity above a threshold weaken a random incoming synapse; neurons with activity below the threshold strengthen one. Starting from either subcritical or supercritical initial weights, repeated application of this rule drives the network toward a critical regime in which the branching parameter sigma approaches one and the largest finite-time Lyapunov exponent lambda0 approaches zero. When combined with stochastic gradient descent, the rule counteracts the training-induced drift toward supercritical dynamics, but it also reduces classification accuracy because its weight updates are misaligned with task gradients.

How it works

The authors model a feed-forward deep network as a discrete-time dynamical system in which layer-to-layer propagation is one step of the dynamics. Each neuron state is a real-valued continuous quantity updated by S(t+1) = g(S(t) W(t)), where g(x) = 1.7159 tanh(0.6666 x) is the scaled hyperbolic tangent used in classic deep multilayer perceptrons. Weights are initialized uniformly in [-0.05, 0.05] and then multiplied by a global scaling factor m that acts as a control parameter for effective coupling strength.

Criticality is quantified with two measures. The branching parameter sigma is the ratio of average output-layer activity to average input-layer activity, where activity of a neuron is the standard deviation of its activation across a batch of inputs. The largest finite-time Lyapunov exponent lambda0 is computed from the maximum singular value of the input-output Jacobian and measures how much infinitesimal perturbations grow. Subcritical networks have sigma less than one and lambda0 less than zero; supercritical networks have sigma greater than one and lambda0 greater than zero; criticality sits at sigma = one and lambda0 = zero.

The homeostatic rule is deliberately minimal. After each forward pass, a random subset of neurons is selected. For each selected neuron, the standard deviation of its activation across the current batch is compared with a threshold T equal to the mean activity of the input layer. If the neuron is active, a randomly chosen incoming connection is weakened by a fixed step c; if inactive, one is strengthened by c. The sign of the weight is preserved, only its magnitude changes.

In experiments on MNIST, the authors initialize networks with m ranging from 0.25, where activity vanishes within a few layers, to 2.0, where activity grows layer by layer. Using nine hidden layers of 1,600 neurons, c = 0.001, batch size 32, and updating 1 percent of incoming links for 50 percent of neurons per iteration, all trajectories converge toward sigma = 1, lambda0 = 0, and a common asymptotic mean coupling strength. The plots show means over five runs with shaded one-standard-deviation regions.

When standard stochastic gradient descent is added, gradient updates alone push the network into the supercritical regime, with lambda0 becoming increasingly positive. Adding the homeostatic rule after each SGD step counteracts this drift in a dose-dependent way: larger values of c keep lambda0 closer to zero. However, classification accuracy falls monotonically with c, and the cosine similarity between SGD and homeostatic weight-update vectors is negative, indicating that the two rules move weights in substantially different directions.

Where a skeptic should push

The most load-bearing assumption is that a binary active-inactive threshold and random single-synapse updates are sufficient to stabilize something as complex as a living neural network. The paper demonstrates convergence only in deep multilayer perceptrons on MNIST, not in recurrent or biological networks. Brain organoids have recurrent connections, neuromodulation, glial cells, and developmental drift, none of which appear in the model.

The trade-off with task performance is also important. Pure SGD achieves the highest test accuracy, and stronger homeostatic adaptation progressively degrades it. That suggests the rule is stabilizing but not learning-friendly. For organoid intelligence, where the goal is usually to elicit useful computation rather than merely prevent seizures, this is a serious limitation.

Finally, criticality is a seductive concept, but its functional role in biological and artificial networks remains debated. The authors cite prior work linking critical dynamics to improved learning and prevention of model collapse, but those associations are correlational. Maintaining sigma = 1 and lambda0 = 0 is not automatically the same as maintaining useful computation.

What this means for organoid stability control

The non-obvious implication is that an organoid interface might keep living neural tissue in a functional regime without solving a global control problem. Current closed-loop systems typically compute some aggregate measure of network activity and then apply a global stimulus or feedback gain. This paper suggests that a local rule, where each electrode or each stimulated site adjusts its drive based only on the variance of the signal it observes, could collectively push the network toward criticality. The mechanism from the source is the negative feedback between local activity and coupling strength: hot neurons weaken their inputs, cold neurons strengthen theirs.

The opportunity is a simpler, more scalable controller for organoid cultures. Brain organoids often show spontaneous epileptiform bursts or silent quiescent periods, either of which degrades their usefulness as a compute substrate. A homeostatic controller could in principle sit on the electrode array, measure activity variability per channel, and nudge stimulation amplitudes or synaptic weights in a closed loop. Because the rule uses only local information, it would not require a central model of the full organoid state, which is attractive when the connectivity map is unknown.

The threat is the same trade-off the authors observe in artificial networks. A controller that keeps the organoid near criticality might also suppress the very plasticity and task-specific adaptation that make biological computing interesting. If homeostatic stabilization and learning are misaligned, then either the organoid remains stable but unlearnable, or it learns but drifts toward runaway activity. There is also a dual-use concern: a controller that can hold a living neural network in a desired dynamical regime could be used to induce pathology as easily as to prevent it.

Ethically, the work sharpens the question of how much external control over a biological neural system is acceptable. A stability controller is defensible as a welfare measure, but the boundary between welfare enhancement and behavioral engineering is not obvious and should be drawn before the technology is deployed.

The bottom line

This is a clean theoretical-and-empirical paper that extends the idea of self-organized criticality from abstract networks and biological models to modern deep neural networks. The local rule is simple, the convergence is robust across initial conditions, and the conflict with gradient-based learning is demonstrated rather than hidden. What remains is to test whether similar local rules can stabilize living neural tissue and whether they can be made compatible with useful plasticity.

For organoid intelligence, the paper is best read as a control-theory hint. It suggests that stability does not necessarily require a global model, but it also warns that stability and learning can pull in opposite directions. The claim would be strengthened by a demonstration in a biological or organoid preparation. It would be weakened if the active-inactive threshold proves too crude for recurrent, heterogeneous living networks.