deep learning 优化方法(未完成待编辑）

SGD(Stochastic Gradient Descent)

ASGD(Averaged Stochastic Gradient Descent)

CG(Conjungate Gradient)

LBFGS(Limited-memory Broyden-Fletcher-Goldfarb-Shanno)

SGD 随机梯度下降

（ref:https://en.wikipedia.org/wiki/Stochastic_gradient_descent）

SGD解决了梯度下降的两个问题：收敛速度慢和陷入局部最优。修正部分是权值更新的方法有些许不同。

Stochastic gradient descent (often shortened in SGD) is a stochastic approximation of the gradient descent optimization method for minimizing an objective function that is written as a sum of differentiable functions

pseudocode

Choose an initial vector of parameters and learning rate $\eta$ .
Repeat until an approximate minimum is obtained:
- $\! w := w - \eta \nabla Q_i(w).$
- Randomly shuffle examples in the training set.
- For $\! i=1, 2, ..., n$ , do:

example

Let's suppose we want to fit a straight line $y = \! w_1 + w_2 x$ to a training set of two-dimensional points $\! (x_1, y_1), \ldots, (x_n, y_n)$ using least squares. The objective function to be minimized is:

$Q(w) = \sum_{i=1}^n Q_i(w) = \sum_{i=1}^n \left(w_1 + w_2 x_i - y_i\right)^2.$

The last line in the above pseudocode for this specific problem will become:

$\begin{bmatrix} w_1 \\ w_2 \end{bmatrix} := \begin{bmatrix} w_1 \\ w_2 \end{bmatrix} - \eta \begin{bmatrix} 2 (w_1 + w_2 x_i - y_i) \\ 2 x_i(w_1 + w_2 x_i - y_i) \end{bmatrix}.$

标准梯度下降和随机梯度下降之间的关键区别

–标准梯度下降是在权值更新前对所有样例汇总误差，而随机梯度下降的权值是通过考查某个训练样例来更新的

–在标准梯度下降中，权值更新的每一步对多个样例求和，需要更多的计算

–标准梯度下降，由于使用真正的梯度，标准梯度下降对于每一次权值更新经常使用比随机梯度下降大的步长

–如果标准误差曲面有多个局部极小值，随机梯度下降有时可能避免陷入这些局部极小值中

梯度下降需要把m个样本全部带入计算，迭代一次计算量为m*n^2

随机梯度下降每次只使用一个样本，迭代一次计算量为n^2，当m很大的时候，随机梯度下降迭代一次的速度要远高于梯度下降

ASGD 平均随机梯度下降

(ref:https://www.quora.com/How-does-Averaged-Stochastic-Gradient-Decent-ASGD-work)

（ref:https://en.wikipedia.org/wiki/Conjugate_gradient_method）

If we choose the conjugate vectors p_k carefully, then we may not need all of them to obtain a good approximation to the solution x_∗. So, we want to regard the conjugate gradient method as an iterative method. This also allows us to approximately solve systems where n is so large that the direct method would take too much time.

We denote the initial guess for x_∗ by x₀. We can assume without loss of generality that x₀ = 0 (otherwise, consider the system Az = b − Ax₀ instead). Starting with x₀ we search for the solution and in each iteration we need a metric to tell us whether we are closer to the solution x_∗ (that is unknown to us). This metric comes from the fact that the solution x_∗ is also the unique minimizer of the following quadratic function; so if f(x) becomes smaller in an iteration it means that we are closer to x_∗.

$f(\mathbf{x}) = \tfrac12 \mathbf{x}^\mathsf{T} \mathbf{A}\mathbf{x} - \mathbf{x}^\mathsf{T} \mathbf{b}, \qquad \mathbf{x}\in\mathbf{R}^n.$

This suggests taking the first basis vector p₀ to be the negative of the gradient of f at x = x₀. The gradient of f equals Ax − b. Starting with a "guessed solution" x₀ (we can always guess x₀ = 0 if we have no reason to guess for anything else), this means we take p₀ = b − Ax₀. The other vectors in the basis will be conjugate to the gradient, hence the name conjugate gradient method.

Let r_k be the residual at the kth step:

$\mathbf{r}_k = \mathbf{b} - \mathbf{Ax}_k.$

Note that r_k is the negative gradient of f at x = x_k, so the gradient descent method would be to move in the direction r_k. Here, we insist that the directions p_k be conjugate to each other. We also require that the next search direction be built out of the current residue and all previous search directions, which is reasonable enough in practice.

The conjugation constraint is an orthonormal-type constraint and hence the algorithm bears resemblance to Gram-Schmidt orthonormalization.

This gives the following expression:

$\mathbf{p}_{k} = \mathbf{r}_{k} - \sum_{i < k}\frac{\mathbf{p}_i^\mathsf{T} \mathbf{A} \mathbf{r}_{k}}{\mathbf{p}_i^\mathsf{T}\mathbf{A} \mathbf{p}_i} \mathbf{p}_i$

(see the picture at the top of the article for the effect of the conjugacy constraint on convergence). Following this direction, the next optimal location is given by

$\mathbf{x}_{k+1} = \mathbf{x}_k + \alpha_k \mathbf{p}_k$

with

$\alpha_{k} = \frac{\mathbf{p}_k^\mathsf{T} \mathbf{b}}{\mathbf{p}_k^\mathsf{T} \mathbf{A} \mathbf{p}_k} = \frac{\mathbf{p}_k^\mathsf{T} (\mathbf{r}_{k-1}+\mathbf{Ax}_{k-1})}{\mathbf{p}_{k}^\mathsf{T} \mathbf{A} \mathbf{p}_{k}} = \frac{\mathbf{p}_{k}^\mathsf{T} \mathbf{r}_{k-1}}{\mathbf{p}_{k}^\mathsf{T} \mathbf{A} \mathbf{p}_{k}},$

where the last equality holds because p_k and x_k-1 are conjugate.

LBFGS

deep learning 优化方法(未完成待编辑）

你可能感兴趣的:(deep learning 优化方法(未完成待编辑）)