Design and develop in C++ a PyTorch extension for a convolutional neural network input layer for sparse genotype data, computing the forward and backward passes with 20-fold reduction in time complexity in large sample sizes.
•
Explored time/space-efficient sparse matrix structures and operations to exploit sparsity in large genetic datasets.
•
Devised novel algorithms that computes the Genomic Relationship Matrix with up to 200-fold reduction in computational complexity and Linkage Disequilibrium score with up to 20-fold reduction, depending on sample size.
•
Developed codebase in C++ for estimating heritability 3x faster than the current leading algorithm.
