Struct vikos::teacher::Nesterov

pub struct Nesterov {
    pub l0: f64,
    pub t: f64,
    pub inertia: f64,
}

Nesterov accelerated gradient descent

Like accelerated gradient descent, Nesterov accelerated gradient descent includes a momentum term. In contrast to regular gradient descent, the acceleration is not calculated with respect to the current position, but to the estimated new one. Source: G. Hinton's lecture 6c

Fields

l0: f64

Start learning rate

t: f64

Smaller t will decrease the learning rate faster

After t events the start learning rate will be a half l0, after two t events the learning rate will be one third l0, and so on.

inertia: f64

To simulate friction, please select a value smaller than 1 (recommended)

Trait Implementations

impl<M> Teacher<M> for Nesterov where M: Model

type Training = (usize, Vec<f64>)

Contains state which changes during the training, but is not part of the expertise

fn new_training(&self, model: &M) -> (usize, Vec<f64>)

Creates an instance holding all mutable state of the algorithm

fn teach_event<Y, C>(&self, training: &mut (usize, Vec<f64>), model: &mut M, cost: &C, features: &M::Input, truth: Y) where C: Cost<Y>, Y: Copy

Changes models coefficents so they minimize the cost function (hopefully)