Score Broadcast and Decorrelation: A General Framework for Broadcast-Based Credit Assignment

Researchers introduce Score Broadcast and Decorrelation (SBD), a theoretical framework that generalizes biologically plausible credit assignment mechanisms across diverse loss functions beyond MSE. The framework unifies error broadcast—an alternative to backpropagation that avoids weight transport—under a single orthogonality principle, with experimental validation showing improvements over existing broadcast approaches on image classification tasks.

This research advances the theoretical understanding of biologically plausible learning algorithms by extending recent work in error broadcast mechanisms. The original Error Broadcast and Decorrelation framework applied only to mean-squared-error losses, limiting its scope. The new SBD framework removes this constraint by identifying a generalizable orthogonality principle: the output score (gradient of loss with respect to final output) remains orthogonal to hidden-layer activations when the optimal score has conditional mean zero. This elegant theoretical insight applies to cross-entropy, Bregman divergences, proper scoring rules, and exponential-family losses—essentially all standard differentiable losses used in modern machine learning.

The work bridges neuroscience and machine learning by grounding the three-factor learning rule—a prominent theory of neuromodulation in biological brains—in mathematical principles applicable to arbitrary losses. The neuromodulatory factor emerges naturally as the broadcast loss score, connecting biological plausibility with mathematical rigor. The introduction of score vector expansion further enriches the broadcast signal while maintaining theoretical guarantees.

Experimentally, SBD demonstrates substantial improvements over existing broadcast approaches on CIFAR-10 and Tiny ImageNet, validating the theoretical contributions. For the AI research community, this work strengthens the case that biologically plausible learning mechanisms can match or exceed conventional backpropagation-based methods while remaining neurologically feasible. The framework's generality positions it as a foundation for future investigation into alternative credit assignment mechanisms, potentially influencing neuromorphic computing and brain-inspired AI architectures.

• →SBD unifies broadcast-based credit assignment across all standard differentiable loss families through a single orthogonality principle.
• →The framework mathematically grounds the three-factor learning rule from neuroscience, connecting biological plausibility to rigorous theory.
• →Score vector expansion technique enriches broadcast signals while preserving the orthogonality framework's theoretical guarantees.
• →Experimental results on CIFAR-10 and Tiny ImageNet show substantial improvements over existing broadcast-based credit assignment methods.
• →The work removes the previous limitation of broadcast mechanisms being restricted to MSE loss, enabling broader applicability.