Deep multi-task mining Calabi-Yau four-folds

Abstract We continue earlier efforts in computing the dimensions of tangent space cohomologies Calabi-Yau manifolds using deep learning. In this paper, we consider dataset all four-folds constructed as complete intersections products projective spaces. Employing neural networks inspired by state-of-the-art computer vision architectures, improve benchmarks and demonstrate that four non-trivial Hodge numbers can be learned at same time a multi-task architecture. With 30 % (80 %) training ratio, reach an accuracy 100 for h(1,1) 97 h(2,1) (100 both), 81 (96 h(3,1), 49 (83 h(2,2). Assuming Euler number is known, it easy to compute, taking into account linear constraint arising from index computations, get total accuracy.