What is the difference between Maximum Likelihood Estimation & Gradient Descent?

What are the pro & cons of both the methods?

• I am not looking for just definition of these two methods which I already have from Google search. I am trying to understand which method is preferred in which case. Eg: For Bigdata will one work better than other etc. I couldn't find any good material which talks about the practical aspects etc. Commented Nov 27, 2015 at 14:42
• How is a raven like a writing desk?
– whuber
Commented Nov 27, 2015 at 15:51
• @ML_Pro GD does not relate anyhow to statistical modeling, it is an algorithm. You probably could start with some introductory statistics handbook to get a better understanding of statistical inference before going into learning the tools (like GD) for solving statistical problems.
– Tim
Commented Nov 27, 2015 at 15:53
• Did you mean to ask the difference between Gradient Descent and Expectation Maximization (which is typically used to solve the optimization problem in MLE)?
– Sobi
Commented Dec 13, 2015 at 0:50

Maximum likelihood estimation is a general approach to estimating parameters in statistical models by maximizing the likelihood function defined as

$$L(\theta|X) = f(X|\theta)$$

that is, the probability of obtaining data $X$ given some value of parameter $\theta$. Knowing the likelihood function for a given problem you can look for such $\theta$ that maximizes the probability of obtaining the data you have. Sometimes we have known estimators, e.g. arithmetic mean is an MLE estimator for $\mu$ parameter for normal distribution, but in other cases you can use different methods that include using optimization algorithms. ML approach does not tell you how to find the optimal value of $\theta$ -- you can simply take guesses and use the likelihood to compare which guess was better -- it just tells you how you can compare if one value of $\theta$ is "more likely" than the other.

Gradient descent is an optimization algorithm. You can use this algorithm to find minimum (or maximum, then it is called gradient ascent) of many different functions. The algorithm does not really care what is the function that it minimizes, it just does what it was asked for. So with using optimization algorithm you have to know somehow how could you tell if one value of the parameter of interest is "better" than the other. You have to provide your algorithm some function to minimize and the algorithm will deal with finding its minimum.

You can obtain maximum likelihood estimates using different methods and using an optimization algorithm is one of them. On another hand, gradient descent can be also used to maximize functions other than likelihood function.

• @ML_Pro I provided two links where you can find detailed information, I don't think there is a need to duplicate these answers.
– Tim
Commented Nov 27, 2015 at 14:34
• @ML_Pro as I wrote in my answer, they are different things and you cannot compare them...
– Tim
Commented Nov 27, 2015 at 14:45
• Yes but MLE is a general approach and GD is just an algorithm you can use to minimize a number of different functions. It is like you compared algebra to pocket calculator...
– Tim
Commented Nov 27, 2015 at 14:56
• MLE specifies the objective function (the likelihood function); GD finds the optimal solution to a problem once the objective function is specified. You can use GD (or other optimization algorithms) to solve a maximum likelihood problem, and the result will be the maximum likelihood estimator. Commented Nov 27, 2015 at 15:58
• @ML_Pro this is described in the links I provided in my answer. In short: yes it is a product of pdf's. Product because we assume that the data is i.i.d. It is defined in terms of pdf's because we are talking about probability model.
– Tim
Commented Dec 6, 2016 at 17:46

Usually, when we get likelihood function $$f = l(\theta)$$, then we solve equation $$\frac{ df }{ d\theta } = 0$$.

we can get the value of $$\theta$$ that can give max or min value of f, done!

But logistic regression's likelihood function no closed-form solution by this way. So we have to use other method, such as gradient descent.

• @Tim, you can see something from here, courses.cs.washington.edu/courses/cse446/13sp/slides/… Commented Dec 1, 2016 at 13:10
• "The regression coefficients are usually estimated using maximum likelihood estimation" (en.wikipedia.org/wiki/Logistic_regression)
– Tim
Commented Dec 1, 2016 at 13:20
• Maximum likelihood estimation do is a kind of method for estimating regression coefficients, but we have several ways to find the solution of MLE. So using likelihood function + gradient descent(which to get the solution of likelihood function) is still a way to do MLE. Commented Dec 1, 2016 at 13:56
• You also can see this sentence Unlike linear regression, we can no longer write down the MLE in closed form. Instead, we need to use an optimization algorithm to compute it. For this, we need to derive the gradient and Hessian. from Machine Learning: a Probabilistic Perspective , Kevin Murphy. Commented Dec 1, 2016 at 13:58
• ...then wording of your answer is confusing as it sounds like you are saying that for logistic regression we are not using ML and instead we use GD.
– Tim
Commented Dec 1, 2016 at 13:59