Structured flexibility in recurrent neural networks via neuromodulation.

Area of Science:

• Theoretical and systems neuroscience
• Computational neuroscience
• Artificial intelligence modeling

Background:

• Understanding biological intelligence relies on computational models.
• Task-optimized recurrent neural networks (RNNs) are widely used but have fixed synaptic weights.
• Biological neural networks feature dynamic synaptic weights modulated by chemicals.

Purpose of the Study:

• To explore the computational impact of synaptic gain scaling, a form of neuromodulation.
• To introduce a neuromodulated RNN (NM-RNN) model with dynamic weight adjustments.
• To investigate how neuromodulation affects RNN performance and mechanisms.

Main Methods:

• Developed a neuromodulated RNN (NM-RNN) model using task-optimized low-rank RNNs.
• Implemented a neuromodulatory subnetwork to dynamically scale recurrent weights.
• Conducted empirical experiments on canonical tasks and theoretical analyses.

Main Results:

• The NM-RNN model demonstrated higher accuracy in training and generalization compared to standard low-rank RNNs.
• Neuromodulatory gain scaling was shown to enable gating mechanisms within the network.
• Analysis revealed how task computations are distributed within the low-rank dynamics of trained NM-RNNs.

Conclusions:

• Dynamic synaptic gain scaling offers structured flexibility, enhancing RNN performance.
• Neuromodulation provides a mechanism for implementing flexible computational control in neural networks.
• The NM-RNN model offers a more biologically plausible and computationally effective approach to modeling intelligence.