Cross-species dendritic computation for biologically inspired decision-making

How can dendritic computation be incorporated into artificial neural networks? Biological evidence suggests that dendrites combine linear weighted integration with nonlinear input modulation to shape neuronal and network-level functions. Motivated by this principle, we constructed species-parameterized working-memory spiking recurrent networks using mouse, macaque, and human anatomical constraints. Within this model, macaque- and human-parameterized networks showed greater temporal persistence and stronger preservation of working-memory effects under stress-like perturbations than did the mouse-parameterized network. We therefore propose a three-compartment nonlinear-parameterized leaky integrate-and-fire model (TNLP-LIF), which combines linear and nonlinear dendritic computations. Applied to distributed reinforcement learning, TNLP-LIF improved integration of encoded and non-encoded information and achieved sustained performance gains on Atari tasks. Analyses of learned integration ratios, low-dimensional firing dynamics, and ablation experiments further support the computational benefit of combining linear and nonlinear dendritic mechanisms in artificial neural networks.