Assume f is a C2-smooth function from
This is a common task usually solved by Gradient descent, and by Newton method if
In 2014 Dauphin et al. introduced and proved a novel method, that they called "Saddle-free Newton" [1]. It is designed to overcome saddle points of arbitrary function.
Though this method is proven powerful both empirically (See figure 4 in chapter 7 in [1]) and theoretically, a public implementation does not exist. This package is an attempt to fill this gap.
References: [1] Yann N. Dauphin et al. (2014) Identifying and attacking the saddle point problem in high-dimensional non-convex optimization, https://arxiv.org/abs/1406.2572