Task classes that require strong coupling in reservoir RNNs

Determine which categories of computational tasks in reservoir computing require or benefit from strongly coupled recurrent neural networks with large recurrent coupling strength, as opposed to tasks that can be performed by weakly coupled recurrent neural networks operating with only mild nonlinearity.

Background

The study shows that strongly coupled reservoirs often suffer from runaway excitation, leading to oscillatory, chaotic, or fixed-point attractors that degrade performance. The authors mitigate these issues via structural heterogeneity (e.g., weak rows) and automatic gain control (AGC), enabling good performance without fine-tuning.

They note prior results identifying tasks that work with mildly nonlinear, weakly coupled RNNs, and suggest that other tasks—such as binary sequence prediction—may benefit from strong coupling where neurons operate in saturation. The precise delineation of which tasks truly need strong coupling remains unresolved.