Datatician
Recent Articles
Implementing Autodiff: First Attempts at Zig Abstractions
This article documents some of the early design decisions made when creating Zigrad. It discusses implementing scalar-valued computation graphs, the trade-offs between simplicity and extensibility, and the challenges of designing abstractions in a language without native interface features. Key lessons learned inform the groundwork for future tensor-valued implementations.
Marco Christiani
Compress a Neural Network with the SVD
Exploring the potential in low-rank approximations for the compression of neural networks and the latest applications of matrix decomposition to LLMs at scale.
Marco Christiani
Copulas
Measures of risk, dependence, and the simulation of market data for statistical arbitrage.
Marco Christiani
Eigen Intuition
Eigenvalues and eigenvectors remain among the most important concepts in the study of feature extraction, behavior of dynamical systems, matrix factorizations, and many more data science topics. A brief overview of these concepts and how they can inform our understanding of matrices follows.
Marco Christiani