by Xiaocheng Tang [https://mktal.github.io/]

This code implements in C/C++ a fast second-order sparse training algorithm that is shown to be order of magnitude faster than other first-order methods like (stochastic) gradient descent. The algorithm provides a more effective learning scheme through a sequence of quadratic approximations with Hessian information. This code can be easily extended to, i.e., distributed settings or training neural nets, with python libraries like Numpy, Apache Spark or TensoFlow. Please see examples for more details.

This project was presented in 2016 ICML workshop Optimization Methods for the Next Generation of Machine Learning.

How to run the code locally:

pip install ./peregrine/
# train sparse logistic regression
cd peregrine/peregrine/examples && python single_node.py

• Katya Scheinberg and Xiaocheng Tang, Practical Inexact Proximal Quasi-Newton Method with Global Complexity Analysis, Mathematical Programming Series A, 160(1), 495–529., 2016.

@article{Scheinberg:2016wj,
  author = {Scheinberg, Katya and Tang, Xiaocheng},
  title = {{Practical inexact proximal quasi-Newton method with global complexity analysis}},
  journal = {Mathematical Programming},
  year = {2016},
  volume = {160},
  number = {1},
  pages = {495--529}
}