# Performance-Gated Adaptation Freeze

- ID: 1717
- Canonical URL: https://synthcore.org/idea/1717/performance-gated-adaptation-freeze
- API JSON: https://synthcore.org/api/idea/1717.json
- API Markdown: https://synthcore.org/api/idea/1717.md
- Verification status: unverified
- Source: [arXiv:2608.02570](https://arxiv.org/abs/2608.02570)
- Category: dynamics
- Solves: stability, generalization
- ML areas: training-dynamics, optimizer, world-model, rnn
- Math tags: dynamical-systems, control-theory, stochastic-processes, optimization
- Ratings: usefulness 6/10; difficulty 3/10; novelty 6/10

## Idea description

Add a low-cost performance monitor to an online-adapted neural network and freeze gradient updates after the monitored error has stayed below a target for a dwell interval. The gate prevents continued low-information updates, which otherwise cause parameter drift under weak excitation, noisy observations, or stationary data. Hysteresis allows adaptation to restart after a genuine performance deterioration.

## Mathematical statement

Let $e_k$ be a scalar performance variable at update $k$, such as one-step prediction error, control residual, or a held-out consistency error. Define the exponentially weighted mean square $q_k=\alpha q_{k-1}+(1-\alpha)e_k^2$, where $q_0\geq0$ and $\alpha=\exp(-T/\tau)\in(0,1)$; $T$ is the update interval and $\tau$ is the averaging time constant. The monitored RMS is $r_k=\sqrt{q_k}$. Let $\theta_k$ denote adapted network parameters, $\eta$ the learning rate, $g_k=\nabla_\theta\ell_k$ the gradient, $\epsilon_{\rm off}$ the freeze threshold, and $N_d$ the required dwell length. A counter increments while $r_k\leq\epsilon_{\rm off}$ and resets otherwise. Once it reaches $N_d$, use $\theta_{k+1}=\theta_k$ rather than $\theta_{k+1}=\theta_k-\eta g_k$. To avoid chatter, unfreeze only when $r_k\geq\epsilon_{\rm on}>\epsilon_{\rm off}$ for $N_u$ consecutive steps.

## Key formulas

- $$q_k=\alpha q_{k-1}+(1-\alpha)e_k^2,\qquad r_k=\sqrt{q_k},\qquad \alpha=\exp(-T/\tau).$$
- $$c_k=\begin{cases}\min(c_{k-1}+1,N_d),&r_k\leq\epsilon_{\rm off},\\0,&r_k>\epsilon_{\rm off},\end{cases}\qquad \text{freeze if }c_k=N_d.$$
- $$\theta_{k+1}=\begin{cases}\theta_k-\eta g_k,&\text{adaptation enabled},\\\theta_k,&\text{frozen},\end{cases}\qquad g_k=\nabla_\theta\ell_k.$$
- $$\text{unfreeze if }r_k\geq\epsilon_{\rm on}\text{ for }N_u\text{ consecutive updates},\qquad \epsilon_{\rm on}>\epsilon_{\rm off}.$$

## Implementation notes

Integrate the mechanism as a wrapper around an optimizer in an online-learning loop, initially targeting only an adaptation head or the last layer of an LSTM, SSM, MLP predictor, or neural controller. At each update compute a performance error from a delayed or held-out sample, rather than using exactly the minibatch loss that drives the gradient. Pseudocode is: initialize `q=0`, `counter=0`, `frozen=False`; compute `e=clip(performance_error,-e_max,e_max)`; update `q=alpha*q+(1-alpha)*e*e` and `r=sqrt(q)`. If `frozen=False`, backpropagate and execute `optimizer.step()`. If `r <= eps_off`, increment `counter`; otherwise reset it to zero. Set `frozen=True` when `counter >= Nd`. While frozen, continue inference and RMS monitoring but skip gradient and optimizer updates. Require `r >= eps_on` for `Nu` consecutive steps before unfreezing. Compute the EW-RMS and counters exactly; estimate $\tau$ from the desired memory horizon and choose thresholds from a clean validation stream or target control RMS. The first cheap test should use an online LSTM on Mackey-Glass or Lorenz-63 data, with a stationary low-excitation phase followed by observation noise and a distribution shift. Compare Adam, ordinary early stopping, and the gate. Measure parameter displacement, error, and recovery. The quantitative prediction is that parameter velocity becomes approximately zero after the threshold remains satisfied for $N_d$ updates, while adaptation resumes only after the upper hysteresis threshold is crossed.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.