# Excitation-Controlled Recurrent Learning

- ID: 2058
- Canonical URL: https://synthcore.org/idea/2058/excitation-controlled-recurrent-learning
- API JSON: https://synthcore.org/api/idea/2058.json
- API Markdown: https://synthcore.org/api/idea/2058.md
- Verification status: mechanism_failed
- Source: [arXiv:2608.09722](https://arxiv.org/abs/2608.09722)
- Category: dynamics
- Solves: sample-efficiency, accuracy, stability
- ML areas: rnn, ssm, world-model, curriculum, training-dynamics
- Math tags: dynamical-systems, control-theory, linear-algebra, optimization, statistics
- Ratings: usefulness 7/10; difficulty 6/10; novelty 8/10

## Idea description

Expose a recurrent model to deliberately designed input pulses or latent-state perturbations instead of training only on passive trajectories. Choose perturbations that maximize the smallest eigenvalue of the accumulated feature Gramian, making otherwise indistinguishable recurrent couplings recoverable and reducing uncertainty in long-horizon predictions.

## Mathematical statement

For hidden dynamics \(\dot{x}=F_{\theta}(x)+Bv(t)\), let \(J(t)=\partial F_{\theta}/\partial x\) along a nominal trajectory and let \(\Psi(t,\tau)\) satisfy \(\partial_t\Psi(t,\tau)=J(t)\Psi(t,\tau)\), \(\Psi(\tau,\tau)=I\). The finite-horizon controllability or excitation Gramian is \(W_T=\int_0^T\Psi(T,\tau)BB^{\top}\Psi(T,\tau)^{\top}d\tau\). Perturbations are informative when \(\lambda_{\min}(W_T)\) is bounded away from zero. For neural identification, use an empirical feature Gram matrix \(G(v)=\Phi(v)^{\top}\Phi(v)/N\), where \(\Phi(v)\) contains trajectory features or parameter sensitivities under control sequence \(v\), and select interventions by maximizing \(\lambda_{\min}(G+\epsilon I)\) or \(\log\det(G+\epsilon I)\).

## Key formulas

- $$\dot{x}(t)=F_{\theta}(x(t))+Bv(t),\qquad \frac{\partial\Psi(t,\tau)}{\partial t}=J(t)\Psi(t,\tau),\quad J(t)=\frac{\partial F_{\theta}}{\partial x}(x(t)).$$
- $$W_T=\int_{0}^{T}\Psi(T,\tau)BB^{\top}\Psi(T,\tau)^{\top}\,d\tau.$$
- $$v^{\star}=\arg\max_{v\in\mathcal{V}}\lambda_{\min}(G(v)+\epsilon I),\qquad G(v)=\frac{1}{N}\Phi(v)^{\top}\Phi(v).$$

## Implementation notes

Integrate this as an active-data-collection controller or curriculum module for an RNN, neural state-space model, or world model. During an initial passive rollout, estimate the hidden-state Jacobian using automatic differentiation and identify weakly excited directions. Inject bounded pulses, colored noise, or channel-specific interventions into observable inputs or selected latent coordinates. For each candidate control sequence \(v^{(m)}\), roll out the current model, compute sensitivities \(\partial x_j/\partial\theta_S\) or candidate regulator features, and construct \(G(v)=\Phi(v)^\top\Phi(v)/N\). Select the candidate maximizing \(\lambda_{\min}(G+\epsilon I)\) subject to \(\|v\|_2\leq c\), collect that trajectory, and train with the integral sparse loss. The paper contributes the identifiability and Gramian criterion; the implementation estimates \(J\), \(\Phi\), and \(G\) with autodiff and minibatch rollouts. First test a 20-dimensional sparse tanh RNN with only five observed coordinates, comparing passive Gaussian inputs with greedy pulse design at equal trajectory budgets. The prediction is that interventions increase \(\lambda_{\min}(G)\) by at least 2x and reduce parameter error; when \(\lambda_{\min}(G)\) is near zero, edge recovery and long-horizon prediction should fail sharply, while designed perturbations move the model above this boundary.

## Verification

- Status: mechanism_failed
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a reproducible linear recurrent identification MVP with empirical feature Gramians, greedy pulse selection, parameter recovery, and fixed-seed sweeps. The mechanism checks passed: Gramian minimum eigenvalue scaled as amplitude squared, vanished at an unexcited channel, and approximately followed b squared scaling. However, under equal input energy the designed pulse set had lower lambda_min than passive excitation (0.05425 vs 0.07168), so the central promised Gramian improvement was not observed; active data nevertheless reduced parameter error.

### Mechanism check

- Verdict: Built a reproducible linear recurrent identification MVP with empirical feature Gramians, greedy pulse selection, parameter recovery, and fixed-seed sweeps. The mechanism checks passed: Gramian minimum eigenvalue scaled as amplitude squared, vanished at an unexcited channel, and approximately followed b squared scaling. However, under equal input energy the designed pulse set had lower lambda_min than passive excitation (0.05425 vs 0.07168), so the central promised Gramian improvement was not observed; active data nevertheless reduced parameter error.
- Confidence: 9/10
- Limitations: This is a 2D linear recurrent toy, not the proposed 20D sparse tanh RNN with five observed coordinates. No online neural autodiff Jacobian estimation, latent interventions, sparse loss, long-horizon prediction, FLOP accounting, or robustness to model misspecification was tested.

## Artifacts

- [experiment.py](https://synthcore.org/code/828/experiment.py)
- [report.md](https://synthcore.org/code/828/report.md)
- [results.json](https://synthcore.org/code/828/results.json)
- [Download all files as ZIP](https://synthcore.org/download/828)

## Disclaimer

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