<ul><li>A new projection-based framework for neural network training has been introduced, focusing on gradient-free and parallel learning.</li><li>The framework uses projection operators and iterative projection algorithms, different from traditional gradient-based optimization methods.</li><li>Training is reformulated as a large-scale feasibility problem, involving finding network parameters and states that satisfy local constraints from elementary operations.</li><li>The framework, named PJAX, supports GPU/TPU acceleration and offers a NumPy-like API, showing promising results in training various architectures on standard benchmarks.</li></ul>

A projection-based framework for gradient-free and parallel learning

Discover more