What method is used in training to prevent overfitting by freezing some weights at various points?

The method that prevents overfitting by freezing some weights at various points is dropout, which is utilized during the training of neural networks. Dropout works by randomly setting a fraction of input units to zero during each forward pass through the network, effectively 'dropping out' certain neurons. This process encourages the model to learn more robust features that are not reliant on any specific set of neurons, reducing the likelihood of overfitting to the training data.

By incorporating dropout, the model is forced to generate multiple independent representations of the data, which helps in generalizing better to unseen data. This approach mitigates the risk of the model becoming too complex and sensitive to noise in the training set.

The other methods listed, such as gradient descent, inefficient batching, and increasing batch size, do not specifically target overfitting through the mechanism of freezing weights or dropping connections randomly during training, making dropout the distinct and correct choice for this purpose.