Training a model is minimizing the cost function. To train a logistic regression model, we often define the cross entropy loss (negative sum of log likelihood) as cost function, and to minimize it, we use (mini-batch stochastic) gradient descent.