Towards Adaptive and Stable Compositional Assemblies of Recurrent Neural Network Modules

Caro, Valerio & Ceni, Andrea & Bacciu, Davide & Gallicchio, Claudio. (2025). Towards Adaptive and Stable Compositional Assemblies of Recurrent Neural Network Modules. 675-680. 10.14428/esann/2025.ES2025-48.

Abstract: Recurrent neural networks (RNNs) are computational models regarded as dynamical systems. Modularity is a key ingredient of complex systems. Thus, the composition of RNN modules provides a simple paradigm for building complex computational models, with the potential to approach the human brain capability. We devise strategies for training RNNs assembled into a larger RNN of RNNs, provided with theoretical guarantees of stability that hold during training for the composed global network. Experiments on pixel-by-pixel image classification benchmarks prove the effectiveness of this approach.