Dimensionless Controls of Plasticity Under Alternating Tasks: From Evolutionary Biology to Continual Learning

Technical Summary: Dimensionless Controls of Plasticity Under Alternating Tasks

Problem Statement
The paper addresses the phenomenon of plasticity—the sustained ability to learn new tasks without catastrophic forgetting—across two distinct domains: evolutionary biology and continual learning in deep neural networks. While recent work by García-Galindo and Ahnert (2025) identified four biological factors controlling plasticity in genotype–phenotype maps (neutral set size, Hamming distance, switching period, and mutation rate) under alternating environments, it remains unclear which of these controls translate to gradient-based learning. The central problem is to determine if a "mechanics of learning" can be established, analogous to physical mechanics, by identifying dimensionless quantities that govern plasticity in deep networks trained on alternating Boolean tasks.

Methodology
The author constructs a minimal deep-learning analogue to the biological setting.

• Experimental Setup: A neural network is trained alternately on two Boolean label sets ($y_1$ and $y_2$) over $N=128$ inputs (7-bit functions). The training switches between tasks every $T$ steps using Stochastic Gradient Descent (SGD).
• Metrics:
  • Accuracy Proxies ($a_j$): Defined as $e^{-\ell_j(t)}$, mapping loss to the unit interval $(0, 1]$.
  • Utopia Distance ($d_U$): The time-averaged Euclidean distance from the trajectory to the "utopia point" $(1,1)$, representing joint mastery.
  • Forgetting Amplitude ($\Delta a_1$): The peak-to-peak swing in performance on the first task when switching to the second.
• Controls: The study varies four factors: task disagreement ($r$, the fraction of differing labels), learning rate ($\eta$), switching period ($T$), and neutral set size (NSS).
• Analysis: The author derives theoretical bounds on forgetting and utopia distance. Empirically, they analyze 9,720 trajectories using Type II Analysis of Variance (ANOVA) to determine the effect sizes ($\omega^2$) of each factor.

Key Contributions

1. Reduction to Dimensionless Controls: The study demonstrates that the four biological factors reduce to two dominant dimensionless controls in the deep learning setting:
   • Task Disagreement ($r$): The fraction of inputs where labels differ.
   • Reach ($\eta T$): The product of the learning rate and the switching period, representing the distance traveled in parameter space during one phase.
   • The author shows that neutral set size (NSS) and complexity, while critical in biology, have negligible effects in this gradient-based setting.
2. Theoretical Bounds:
   • Theorem 1 (Forgetting): Derives an upper bound on forgetting amplitude: $\Delta a_1 \lesssim e^{r \eta T} M - 1$, where $M$ is a bound on the forcing term. This establishes that forgetting is jointly controlled by $r$ and $\eta T$.
   • Theorem 2 (Geometric Floor): Proves that $r$ alone fixes a tight lower bound on the utopia distance: $d_U \geq \sqrt{2}(1 - 2^{-r})$. This indicates that perfect joint mastery is geometrically impossible if tasks disagree.
3. Optimal Scaling Law: Empirical results reveal that the optimal reach ($\eta T^*$) follows an approximate inverse power law with respect to task disagreement: $\eta T^* \propto r^{-1.18}$. This suggests a heuristic where the optimal reach can be set based solely on the task disagreement.

Results

• Dominance of Dynamic Factors: ANOVA results confirm that $r$, $\eta$, and $T$ dominate the variance in both $d_U$ and $\Delta a_1$. Specifically, $r$ explains 94.22% of the variance in $d_U$, while $\eta$ and $T$ are the primary drivers of forgetting.
• Negligibility of Neutral Sets: Despite statistical significance due to large sample sizes, the effect size of neutral set size ($\nu$) and complexity ($\tilde{K}$) on plasticity outcomes is negligible ($\omega^2 \approx 10^{-4}$), contrasting sharply with biological findings where NSS is a primary controller.
• Optimal Reach: The heatmap of $d_U$ over $(\eta, T)$ space shows that optimal plasticity occurs along a diagonal where $\eta T$ is constant. As $r$ increases, the optimal $\eta T$ decreases, and the system becomes more sensitive to overfitting.
• Robustness: The findings hold across different optimizers (SGD, Adam, momentum), network widths/depths, and held-out continuous-input tasks, suggesting the controls are not artifacts of the specific Boolean setup.

Significance and Claims
The paper claims to contribute a specific entry to the emerging "mechanics of learning," a framework aiming to organize learning theory around canonical solvable models and dimensionless quantities, similar to the Reynolds number in fluid dynamics.

• Dynamical vs. Geometric Analogy: The primary analogy that survives the translation from biology to deep learning is dynamical rather than geometric. While biological plasticity relies heavily on the geometry of neutral sets (static landscape properties), plasticity in gradient descent is governed by the dynamics of the system (how far the network relaxes within a switching period).
• Design Heuristic: The work provides a practical heuristic for continual learning: the optimal reach (the product of learning rate and switching period) can be estimated from task disagreement alone ($\eta T \approx 1/r$), without needing to probe the loss landscape or network architecture.
• Interdisciplinary Link: By identifying the reach $\eta T$ as a Damköhler number (a ratio of residence time to reaction time in chemical engineering), the paper positions single-trajectory learning alongside driven systems in physics and engineering, suggesting a potential substrate universality in learning mechanics.

The author remains modest regarding the generality of these findings, noting that while the setting is a candidate minimal model for any computational data (as it can be written as a Boolean function), the specific power law exponent and the exact role of neutral sets in ensemble methods or population-based training require further investigation.