Quiz: CC&ET: Machine Learning – Nuts & Bolts: How Does DL Work?

What is the key characteristic feature of deep learning?
- The application of machine learning to medical imaging problems.
- The use of neural network architectures with large numbers of layers.
- The use of specialized numerical optimization methods in machine learning.
- The use of specific software tools to solve machine learning problems.
- The use of neural network architectures with large numbers of parameters.
- All of the above.
Jump to the relevant part of the video for the above question:
What is the definition of learning in the context of deep learning?
- The learning process to identify the best network architecture for a certain problem using cross validation.
- The data labeling process in a deep learning application.
- The identification of neural network parameters such that the output of the network approximates class labels by means of minimizing an error functional.
- All of the above.
Jump to the relevant part of the video for the above question:
It has been demonstrated that with novel, more complex looking network architectures, the difficulty of optimization problems (defined by their degree of non-convexity) encountered in deep learning...
- Decreases, which to some degree explains the success of novel architectures.
- Stays the same. It does not depend on the architecture, only on the particular data that is used for training.
- Increases, which to some degree explains that while these architectures show great performance, they also require a lot more data.
- All of the above.
Jump to the relevant part of the video for the above question:
What is the relationship between training deep neural networks and spline fitting, and what are the consequences?
- There are no relationships, spline fitting is used for curve fitting and deep learning is used to process images.
- According to the representer theorem, training a deep neural network is identical to fitting a high dimensional spline to data, which allows some insight into situations when deep learning generalizes successfully and when it fails.
- According to the sparbuchdrucker theorem, the output layers of deep learning architectures can be substantially simplified with spline fitting methods, which allows a substantial reduction in computational cost.
- All of the above.
Jump to the relevant part of the video for the above question: