Which concept is used to adjust model weights via the backpropagation during training?

The concept that is used to adjust model weights via backpropagation during training is the gradient. In the context of neural networks, backpropagation is a key algorithm that allows the model to learn from the errors it makes. This process involves calculating the gradient of the loss function with respect to the model's weights. The gradient essentially indicates the direction and magnitude of the change needed in the weights to minimize the loss. By using this information, the weights are updated to improve the model's performance.

The other options, while relevant to machine learning, do not specifically pertain to the process of adjusting weights during training. Dropout is a regularization technique aimed at preventing overfitting by randomly setting a portion of neurons to zero during training. Hyperparameter tuning involves optimizing parameters that are not learned directly by the model, such as the learning rate or batch size, rather than adjusting the model's weights directly. Inference optimization focuses on improving the efficiency and speed of model predictions once the model is trained, rather than during the training phase itself.