When you update a neural network's weights during training, which special TensorFlow item *changes its valuedue to automatic differentiation?

When you update a neural network's weights during training, the special TensorFlow item that changes its value due to automatic differentiation is a `tf.Variable`. A `tf.Variable` is a specific type of TensorFlow object designed to represent mutable, stateful tensors. Unlike a standard `tf.Tensor`, which is immutable once created, a `tf.Variable` can have its value updated. In the context of neural networks, `tf.Variable`s are used to hold the learnable parameters of the model, such as the weights and biases. During the training process, TensorFlow's automatic differentiation mechanism, typically implemented using `tf.GradientTape`, tracks the operations performed on these `tf.Variable`s to calculate the gradients of the loss function with respect to their values. Gradients indicate the direction and magnitude of the steepest ascent of the loss function. Once these gradients are computed, an optimizer (e.g., Adam, SGD) takes these gradients and applies them to update the internal numerical value of the `tf.Variable`. This update step directly modifies the `tf.Variable`'s value, moving it in a direction that minimizes the loss, thereby adjusting the network's weights and biases to improve its performance.