MesoNet: perturbation and Jacobian notes
perturb one weight, follow what changes. specify the state + scale of the perturbation before staring at the Jacobian
approach: perturb synaptic weights and examine the resulting change in neuron/network outputs. The journal explores linear stability and Jacobian-based sensitivity to guide update magnitudes.
recorded observation: highly sensitive configurations appeared to generalize poorly; adjusted updates were described as improving convergence. The dated entry does not attach a numerical comparison, split definition or full derivation, so these remain qualitative notes.
References saved in this journal section
Lyapunov heatmap
Eigenvalue distribution
- review questions: what state vector does the Jacobian act on—voltages, traces, weights, or all of them? Is the map continuous-time or one simulation step? How are threshold/reset discontinuities treated? Repeat perturbations at more than one magnitude before assuming a linear approximation is adequate.
- artifact to recover: the actual matrix construction and run output, not only a final eigenvalue plot.
Bifurcation and Lyapunov notes