Sparse Predictive Hierarchies

Testing biologically inspired learning against the strategic depths of chess

Testing biologically inspired learning against chess reveals a clear gap between ambition and results. A 44.9M parameter Sparse Predictive Hierarchy, built on local learning rules, failed catastrophically against a tiny 3.8M CNN trained with standard backpropagation. The extreme ~99% sparsity choked tactical information flow, proving that local updates struggle with global credit assignment in dense, long-range domains. This negative result is valuable, it sharpens the question of whether biological learning can ever scale to tasks requiring strict global feedback.