Quiz: CC&ET: Machine Learning – How to Jump Start Your Deep Learning Research

What are key features one should look for in a deep learning framework?
- Data structures for computational graphs, GPU support.
- High level functions for architecture design, data handling support.
- Automatic calculations of gradients, efficient numerical optimizers.
- All of the above.
Jump to the relevant part of the video for the above question:
To validate the generalizability of a trained model, it is good practice to use the error rate on...
- The data that was used to train the model. If this is as low as possible, it guarantees that the best model was identified by the numerical optimization procedure.
- A validation set that was not used during the actual training of the model, but can be used to design the model architecture and training hyperparameters. This guarantees that the best architecture is selected for this type of problem.
- An independent test set, that is only used once at the end of all developments and that was not available during any part of the development. Only this guarantees that the trained model can truly generalize to new unseen data.
- All of the above.
Jump to the relevant part of the video for the above question:
If you encounter high accuracy on the training set and low accuracy on the validation set, what does this indicate and what is the proper procedure going forward?
- The model does not have enough computational complexity, so you need to increase the size of the model and therefore the number of free parameters.
- The model complexity is too high in relation to the available training data and overfitting occurs. You should decrease the number of parameters in the model, change the model architecture or obtain additional training data.
- The training has not converged properly. You should continue training the same model with the same data.
- All of the above.
Jump to the relevant part of the video for the above question:
In comparison to fully connected layers, convolutional layers usually have...
- Less free parameters because parameters are re-used by exploiting a certain structure (for example translational invariance in the case of images) in the data.
- More free parameters because convolutions are computationally expensive operations.
- The same number of free parameters, because this depends only on the input data.
- All of the above.
Jump to the relevant part of the video for the above question: