Common Optimisers

$E$ is the loss function and $w$ is the model parameters;

Stochastic Gradient Descent

$$\begin{aligned} w_{t+1}&= w_t - \alpha \nabla E(w_t)\end{aligned}$$

SGD with Momentum

Use gradient to update velocity/direction of a particle instead of only updating its position

$$\begin{aligned} m_{t+1} &= \eta m_t + \alpha \nabla E(w_t) \\ w_{t+1}&= w_t - m_{t+1} \end{aligned}$$ This results in equivalent single update as $$w_{t+1}= w_t - \alpha \nabla E(w_t) - \eta m_{t}$$

$\eta$ is the exponential decay factor in $[0,1]$ which determines contribution of previous gradients to the weight change.

Nesterov Accelerated Gradient

The observation behind nesterov momentum is that we will update the parameters by the momentum term anyway, why not calculate gradient at the updated step instead?

$$\begin{aligned} m_{t+1} &= \eta m_t + \alpha \boldsymbol{\nabla E(w_t-\eta m_t)} \\ w_{t+1}&= w_t - m_{t+1} \end{aligned}$$

Since we calculate the gradient at the new location, if there is a difference in direction, the update will be able to correct for the difference. It increases responsiveness of the optimiser.

RMSProp

Root Mean Square Propagation. Divide learning rate of each weight by running average of magnitudes of recent gradients for that weight. $$\begin{aligned} v_{t+1} &= \gamma v_{t} + (1-\gamma)(\nabla E(w_t))^2 \\ w_{t+1}&=w_{t} - \frac{\alpha}{\sqrt{v_{t+1}}} \nabla E(w_t) \end{aligned}$$

Note: $\nabla E(w_t)^2=|\nabla E(w_t)|_F^2$

Adam

Adaptive Moment Estimation is an update to RMSProp. It uses a running average for the gradient as well $$\begin{aligned} v_{t+1} &= \beta_1 v_{t} + (1-\beta_1)(\nabla E(w_t))^2 \\ m_{t+1} &= \beta_2 m_{t} + (1-\beta_2)\nabla E(w_t) \\ m &= \frac{m_{t+1}}{1-\beta_2} \quad v = \frac{v_{t+1}}{1-\beta_1} \\ w_{t+1}&=w_{t} - \alpha \frac{m}{\sqrt{v}+\epsilon} \end{aligned}$$




tags: machine-learning deep-learning