Coherence from Context - Two-Point Neuron Models for Contextual Integration in Visual Information Processing

The integration of contextual information is crucial for coherent perception and cognition. The morphology and conductance properties of cortical pyramidal cells suggest that they operate as “two-point neurons” (TPNs), asymmetrically combining driving sensory input at basal dendrites with modulating context at apical compartments. We present a mechanistic computational TPN model that captures the causal cell-internal apical-basal integration. The model is extended incorporating the interactions between pyramidal cells and local inhibitory interneuron circuits of PV, SOM, and VIP cells. The model rests on guiding principles of asymmetric feedforward-feedback integration, contextual feedback, and pooled inhibition to implement local competition and global cooperation supporting the selective amplification of coherent signals. We validate our approach against detailed multi-compartment pyramidal cell simulations reproducing key electrophysiological phenomena. We then extend it to interacting TPN populations with joint spatial and feature selectivity. In such networks, contextual signals propagate through structured lateral recurrence and top-down feedback, exhibiting contextual integration, coherence formation, and evidence propagation. To support larger-scale network simulations, we derive a reduced mathematical model that preserves the core computational principles of TPNs, while substantially reducing complexity. We demonstrate the model’s applicability in biological vision, showing how it explains motion integration and incremental grouping—processes requiring dynamic resolution of perceptual ambiguity. Finally, we discuss how the proposed framework connects cellular and circuit-level mechanisms of pyramidal neurons to broader questions about cortical computation, the formation of representations for globally consistent perceptual states, and the potential for embedding TPN principles into artificial neural network architectures.