2024 wrap up

[AlbertDominguez]

Issues/stuff to improve

• Much more time than scheduled (it was OK as the rest of the afternoon was free anyway)
• A lot of time spent on the XOR
• Live loss curves plotting didn't show on the exercise, but in the solutions
• IMO the introduction to torch itself was a bit too abrupt by starting with nn.Module directly
• Some people were a bit confused by the loss reduction parameter

Ideas for next year

• Introduce torch as a framework, step-by-step from the beginning: torch.Tensor, ops on it, etc. then nn.Module (layers, combining them, etc.)
• Implement the perceptron in torch (instead of numpy) as a first torch-based task