Which gradient descent method uses the entire dataset to calculate gradients?

Batch gradient descent is the method that utilizes the entire dataset to compute the gradients used for updating the model parameters. By calculating the gradients over the complete dataset, Batch gradient descent ensures that each update is based on comprehensive information, which can lead to a more stable convergence towards the minimum of the loss function. This is especially beneficial when dealing with complex models or when the dataset is not subject to high variability.

In contrast, Mini-batch gradient descent processes the data in small batches, providing a trade-off between the efficiency of Batch and the randomness of Stochastic gradient descent. Stochastic gradient descent updates the model parameters using only a single example at a time, which can introduce considerable noise but allows faster updates. The method known as Stochastic average gradient is related but specifically focuses on averaging gradients over many iterations of Stochastic gradient descent.

Therefore, Batch gradient descent is distinguished by its utilization of the entire dataset for a single gradient calculation, making it the appropriate choice in this context.