Score Broadcast and Decorrelation: A General Framework for Broadcast-Based Credit Assignment
Researchers have developed a new framework called Score Broadcast and Decorrelation (SBD) for credit assignment in neural networks. This framework is designed to work with various differentiable loss functions, offering a biologically plausible alternative to backpropagation. SBD is grounded in an orthogonality principle between the output score and hidden-layer activations, unifying broadcast-based credit assignment across common loss families. Experiments on image datasets demonstrated SBD's effectiveness, with an added score vector expansion technique yielding further improvements. AI
IMPACT Introduces a new theoretical framework for credit assignment that could lead to more efficient and biologically plausible learning algorithms.