Non-negative connectivity causes bow-tie architecture in neural circuits.

Bow-tie architecture (BTA) is widely observed in biological neural systems, yet the underlying mechanism driving its spontaneous emergence remains unclear. In this study, we identify a novel formation mechanism by training multi-layer neural networks under biologically inspired non-negative connectivity constraints across diverse classification tasks. We show that non-negative weights reshape network dynamics by amplifying back-propagated error signals and suppressing hidden-layer activity, leading to the self-organization of BTA without pre-defined architecture. To our knowledge, this is the first demonstration that non-negativity alone can induce BTA formation. The resulting architecture confers distinct functional advantages, including lower wiring cost, robustness to scaling, and task generalizability, highlighting both its computational efficiency and biological relevance. Our findings offer a mechanistic account of BTA emergence and bridge biological structure with artificial learning principles.