On the stability of learning

John J. VastolaarXiv 2026 · arXiv

TL;DR

Learning isn't just about getting the right signal; it's about getting it often enough. This paper demonstrates that if gradient updates are too sparse or intermittent—a common reality in biological brains—learning can become unstable and weight variance can explode, even if the updates are correct on average.

What Problem Does This Paper Address?

In standard machine learning, we usually assume that informative gradients are always available to update weights. However, biological neural networks face "locality constraints": a single synapse doesn't have a global view of the network's error. Instead, it likely receives noisy, sparse, and intermittent feedback.

The authors tackle a fundamental question: Can learning remain stable when feedback is infrequent? While previous research has focused on how the brain might calculate these gradients (credit assignment), this paper focuses on the reliability of the communication. The core challenge is that even if a learning rule is unbiased (correct on average), the high variance introduced by sparse arrivals can lead to a total breakdown of the learning process.

Methods

The author, John J. Vastola, uses analytically tractable models to derive explicit stability conditions. The key methodological steps include:

  • Poisson Gradient Arrivals: Modeling sparse updates by assuming gradients arrive at each synapse according to a Poisson process with a rate $\rho$.
  • Weight Decay (Regularization): Incorporating a continuous relaxation term ($\gamma$) that pulls weights toward zero between gradient arrivals, mimicking biological processes like dendritic spine turnover.
  • Covariance Analysis: Instead of just looking at the average weight (mean dynamics), the author analyzes the evolution of the weight covariance matrix. Learning is defined as "stable" only if the variance of the weights remains finite over time.
  • Mitigation Strategies: The paper mathematically models three biological "fixes":
  • Regularization: Increasing the rate of "forgetting."
    1. Parameter Pooling: Sharing weight information between neighboring neurons (modeled as diffusion).
  • Gradient Averaging: Using "eligibility traces" to average recent signals before applying them.

Datasets

As a theoretical paper, it does not rely on standard large-scale ML datasets like ImageNet. Instead, it uses:

  • Synthetic Linear Models: Single-layer and multi-layer linear networks where the "ground truth" is known.
  • Teacher-Student Setup: A student network is trained to reproduce the output of a teacher network.
  • Nonlinear Validation: A simple two-layer nonlinear network was used to confirm that the insights from the linear theory hold in more complex settings.

Key Insights and Ideas

  • Unbiased is Not Enough: A learning rule can be perfectly accurate on average but still fail if the updates are too variable. This adds a new constraint to what makes a learning model "biologically plausible."
  • The Critical Arrival Rate ($\rho_c$): The author identifies a hard threshold for feedback frequency. If gradients arrive slower than this critical rate, the variance of the weights diverges to infinity.
  • Multiplicative Noise is the Culprit: The instability isn't caused by simple additive noise (like jittery sensors). It stems from the sparse, multiplicative nature of intermittent arrivals, which can cause the system to "overshoot" its target dramatically.
Figure 1: Sparse gradient arrivals: motivation and intuition. (a) A simple example of sparse gradient arrivals: a weight receives intermittent updates, in between which the regularization term produces relaxation toward zero. Three example stochastic simulations are shown. (b) Error as a function of number of arrivals in a toy model. If s/ρ < 1, error always decreases (left); if 1 < s/ρ < 2, error alternates in sign but decreases in magnitude (center); and if s/ρ > 2, learning becomes unstable, and error both alternates in sign and increases in magnitude (right).
  • Forgetting as a Feature: In this framework, weight decay (regularization) isn't just for preventing overfitting; it is essential for stability. By pulling weights back toward a baseline, it prevents the variance from spiraling out of control.

Key Results

  • Stability Condition: For a simple linear model, the critical arrival rate is $\rho > \rho_c := \frac{s^2}{2(s + \gamma)}$, where $s$ is the input variance and $\gamma$ is the regularization strength.
  • Mean vs. Variance: As shown in Figure 2, the average weight might converge perfectly to the target, but if $\rho < \rho_c$, the variance around that average grows without bound.
Figure 2: Mean dynamics (a) look stable regardless of rate, but variance (b, c) explodes if the arrival rate $\rho$ falls below the critical threshold $\rho_c$.
  • Depth is Benign: Surprisingly, increasing the number of layers in a deep linear network does not significantly worsen the stability problem. The critical arrival rate remains relatively stable as depth increases, suggesting that deep biological architectures aren't inherently more prone to this specific type of instability.
  • Pooling and Averaging Work: Both spatial pooling (sharing info with neighbors) and temporal averaging (eligibility traces) successfully lower the $\rho_c$ threshold, making learning possible even with very sparse feedback.

Limitations and Future Work

  • Linearity: The bulk of the formal proofs rely on linear or "near-fixed-point" linear approximations. While the author shows the effect persists in a two-layer nonlinear network, behavior far from the fixed point remains unexplored.
  • Independent Arrivals: The model assumes gradients arrive independently at each synapse. In a real brain, feedback signals are likely correlated across neurons (e.g., a global neuromodulatory pulse), which might change the stability dynamics.
  • Biological Evidence: While the proposed mitigation strategies (like parameter pooling) are theoretically sound, the author notes that direct biological evidence for some of them (like sharing weight info via gap junctions for the purpose of learning stability) is still speculative.

Summary updated

Abstract

Learning algorithms are usually analyzed under the assumption that informative gradients are continuously available. But this is not always true---learning signals in biological neural networks, for example, may be sparse, local, and only intermittently available at any given synapse. In this work, we show that if gradient information is noisy and only intermittently available, learning may be unstable. In particular, even if average learning dynamics converge to the correct optimum, variance in learning may be asymptotically infinite. We explore this point in a number of analytically tractable settings, including single-layer and multi-layer linear networks, and derive explicit conditions under which learning is stable. These conditions complement, but are conceptually distinct from, existing conditions related to the stability of stochastic gradient descent. Our work demonstrates that unbiased credit assignment is not enough for reliable learning; sufficiently frequent and informative feedback is also required, which yields an additional constraint on models of biologically plausible learning.
BibTeX
@misc{vastola2026stability,
  author = {John J. Vastola},
  title = {{On the stability of learning}},
  year = {2026},
  eprint = {2610.35856},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35856},
}
Figure