During which process do updates of model weights occur without waiting for other batches?

Updates of model weights that occur without waiting for other batches are characteristic of asynchronous updates. In this process, each worker can update the model independently and as soon as it has computed its gradients. This allows for faster convergence since updates can be applied immediately, rather than having to wait for a complete cycle of batch processing.

In contrast, synchronous updates require all workers to compute their gradients before any of them can update the model weights. This can lead to delays as the system waits for the slowest worker to finish its computation, which could slow down the training process.

The penalization mechanism and objective function are concepts related to regularization and the goal of model training, respectively, but they do not pertain directly to the timing or methodology of weight updates. Therefore, asynchronous updates are specifically designed for efficiency in the model training process, making this the correct choice.