Structural Correlates of Plasticity Loss

Updated 25 June 2026

Plasticity loss is defined by measurable geometric, topological, and algebraic features that indicate a system's diminishing capacity to adapt.
Quantitative frameworks in neural networks, granular materials, and complex systems reveal that metrics like Hessian rank and defect density reliably signal plasticity loss.
Interventions such as regularization, dynamic sparsification, and noise injection can restore structural markers and help recover adaptive capabilities.

Structural correlates of plasticity loss describe the concrete geometric, topological, and algebraic features of a system’s architecture or state that predict or mediate its diminishing capacity to adapt under continued exposure to new inputs, tasks, or stressors. Plasticity loss is manifest both in physical systems (e.g., yielding in amorphous solids, granular materials) and in artificial systems (e.g., neural networks under continual learning). The identification of structural correlates enables both mechanistic understanding and the rational design of interventions to preserve or restore adaptability.

1. Quantitative Frameworks for Structural Plasticity Loss

The structural basis of plasticity loss has been formalized through different quantitative constructs across domains. In artificial neural networks, plasticity is operationalized via functional metrics such as the effective rank of the Hessian of the loss function, the diversity of feature representations, or the fraction of saturated (“dead”) units (Lewandowski et al., 2023, He et al., 26 Sep 2025, Bonifazi et al., 2024). In complex networks, plasticity is quantified as the ratio of system size to connectivity strength, i.e., $P = N / \sum_{e \in E} |w_e|$ where $N$ is the number of elements and $w_e$ edge weights (Branchi, 26 Mar 2026). In disordered solids and granular packings, local structural disorder or “softness,” and the density of topological defects (e.g., highly distorted coplanar tetrahedra with shape parameter $\delta > 0.245$ ), are the direct correlates of plastic yielding (Cao et al., 2018, Zhang et al., 2022).

These abstractions share the principle that there exists a measurable structural attribute—be it spectral (e.g., rank, eigenvalue density), topological (e.g., defect density, connectivity patterns), or information-theoretic (e.g., feature/covariance rank)—whose degradation or extreme value signals imminent or ongoing loss of adaptation.

2. Hessian Spectral Collapse and Loss-Landscape Geometry

A unifying structural correlate in deep learning is the collapse of the Hessian spectrum associated with the loss function (Lewandowski et al., 2023, He et al., 26 Sep 2025, Lyle et al., 2023). Let $H(\theta) = \nabla^2_\theta L(\theta)$ be the loss Hessian. The effective number of positive eigenvalues (“curvature directions”) provides the manifold’s rank on which gradient descent can act. Empirically, continual learning or prolonged optimization often drives $H$ to be low-rank, i.e., $\mathrm{rank}_\epsilon(H) \to 0$ , a phenomenon termed “spectral collapse.” In this regime, gradient flow is confined to a highly restricted subspace or even trapped in “loss-of-plasticity (LoP) manifolds” from which learning is ineffective (Joudaki et al., 30 Sep 2025).

Feature-space analogues include loss of effective feature rank: a network whose penultimate activations lose diversity (via neural collapse or representational redundancy) is structurally unable to distinguish among new classes or input patterns (Bonifazi et al., 2024).

Hessian spectral collapse is thus both a necessary and sufficient condition for plasticity loss in gradient-based learning: empirical sharp decline in $\mathrm{rank}_\epsilon(H)$ matches the rise in error or inability to learn new tasks (Lewandowski et al., 2023, He et al., 26 Sep 2025).

3. Local Structural Defects in Physical and Amorphous Systems

In physical systems, plasticity loss is governed by the evolution and distribution of local structural indicators. In sheared granular materials, the density of highly distorted coplanar tetrahedra, quantified via a Delaunay shape parameter $\delta = e_\mathrm{max} - 1$ ( $\delta > 0.245$ ), predicts the core sites of plastic rearrangement (Cao et al., 2018). Plastic events (“flips”) are discrete topological transitions—on the Delaunay network, neighbor-switching events correspond to the creation/rotation of 4-ring disclinations. The spatial clustering of such defects under shear leads to macroscopic localization of plastic flow (“shear bands”).

In amorphous solids, the machine-learned scalar “softness,” constructed from local particle descriptors, tightly correlates with local yield strain and future rearrangement propensity (Zhang et al., 2022). Softness field heterogeneity thus is a direct structural predictor of the spatiotemporal sequence of plastic events.

Key structural variables:

System type	Structural correlate	Metric/formula
Deep neural net	Hessian curvature directions	Effective Hessian rank, spectrum: $N$ 0
Granular/amorphous solid	Defect density, softness	Tetrahedron shape $N$ 1, softness $N$ 2
Network system	Connectivity-based plasticity	$N$ 3

4. Mechanisms: Manifolds, Unit Saturation, and Redundancy

Mathematically, loss of plasticity can be characterized as gradient flow becoming confined to invariant manifolds in parameter space—LoP manifolds—either by saturation (frozen units), causing gradients to vanish along certain coordinates, or by excessive symmetry (cloned units), generating representational redundancy (Joudaki et al., 30 Sep 2025). The former arises when units’ activation functions saturate ( $N$ 4), locking incoming weights. The latter occurs when units or blocks become exact linear combinations (clones), so that no gradient in the current architecture can distinguish or optimize their contribution.

A broad simplicity bias in deep learning—where low-rank representations and neural collapse facilitate generalization for static tasks—directly sets up the trap for LoP in highly non-stationary or continual learning, indicating a structural generalization-plasticity tradeoff (Joudaki et al., 30 Sep 2025, Bonifazi et al., 2024).

5. Signatures in Deep Reinforcement Learning and Multi-Task Networks

In deep RL and multi-task networks, several convergent structural signals predict plasticity loss (Klein et al., 2024, Todorov et al., 9 Aug 2025):

Dormant/saturated units: Fraction of units with near-zero activation or in saturated regimes increases (dormancy rate $N$ 5).
Representational collapse: Effective feature rank of the shared feature matrix falls sharply.
Gradient structure: Empirical gradient covariance becomes low-rank or block-structured, indicative of collinear gradients and impaired discrimination.
Parameter norm/spectral explosion: Unbounded growth in $N$ 6 or spectral norm $N$ 7 correlates with plasticity loss.
Loss landscape sharpness: Growth in maximal Hessian eigenvalues (sharp minima).

Practical interventions—dynamic sparsification, feature-rank regularizers, weight resets, or spectral normalization—restore these structural metrics and thereby recover plasticity (Todorov et al., 9 Aug 2025, Lewandowski et al., 2023, Zhang et al., 2022).

6. Structural Interventions and Restoration of Plasticity

Mitigation strategies that act on the structural correlates have been empirically validated:

Regularization of curvature: Imposing L2 weight decay or distributional Wasserstein penalties preserves Hessian rank and prevents collapse (Lewandowski et al., 2023, He et al., 26 Sep 2025).
Feature diversity constraints: Penalties on singular value spectrum or feature covariance effective rank sustain adaptive capacity (He et al., 26 Sep 2025).
Architectural/parameterization design: BatchNorm, LayerNorm, skip/residual connections, or CReLU/PELU activations preserve gradient flow and prevent saturation/frozen units (Lyle et al., 2023, Klein et al., 2024).
Dynamic capacity management: Progressive pruning or sparse evolutionary training maintain high representational diversity and low dormancy (Todorov et al., 9 Aug 2025).
Stochasticity/noise injection: Dropout, Noisy SGD, and continual replacement (CBP) break trapping manifolds in parameter space (Joudaki et al., 30 Sep 2025).

These methods, by directly manipulating the underlying structure, empirically slow or reverse plasticity loss across diverse benchmarks.

7. Broader Structural Principles and Cross-Domain Implications

In the network-based framework, plasticity loss is a robust consequence of deviation from criticality: both excessive connectivity (rigid regime) and insufficient connectivity (unstable regime) cause the system-level plasticity index $N$ 8 to collapse, precluding adaptive response (Branchi, 26 Mar 2026). This principle recapitulates in biological, psychological, social, and ecological networks. Similarly, the formation of local structural defects or clusters (e.g., in solids) marks the threshold for macroscopic yielding.

These structural correlates are thus universal markers transcending model domain, unifying the microscopic (local defects, feature diversity) with the macroscopic (adaptation, learning capacity) in the dynamics of complex systems.

Key references: (Cao et al., 2018, Lewandowski et al., 2023, He et al., 26 Sep 2025, Joudaki et al., 30 Sep 2025, Zhang et al., 2022, Klein et al., 2024, Todorov et al., 9 Aug 2025, Bonifazi et al., 2024, Branchi, 26 Mar 2026, Lyle et al., 2023, Lillo et al., 14 May 2026, Giannini et al., 2021).