In the last video for week 2 “whether to use end-to-end deep learning”, Andrew says that one of the pros is that there’s less hand-designing of components needed. I’m not really sure what this means? Any examples would be really helpful, thank you!
because in end to end deep learning, the deep neural network learns to extract relevant features directly from the raw data. This is contrasting to the traditional pipeline of first providing the raw data , then assembling and/or scaing to providing feature inputs to getting the desired output.
The model steps taken from raw data -cleaning-augmention-algorithm-output is hand designing part of traditional machine learning.
A very common example (which Andrew discussed in the same week for video heading what is end-to-end deep learning) mentioned speech recognition where audio to transcript or text model can use end to end deep learning in case it has more larger data(1000000) but end-to-end requiring more training time where as in the hand-design or traditional machine learning you could work upon on smaller data (3000h audio) with each steps.
This approach could be used based on data, time consumption, computation memory and other factors.
Regards
DP
Another phrase used to describe what Professor Ng is talking about there is “hand-engineered features”. That’s an approach that was more commonly used in the “old days” before Deep Learning basically took over, but it still may have some valid use cases even in 2025. I just tried this search in the new AI powered Google search:
“What are examples of hand-engineered features in Machine Learning”
and I got a nice presentation with some good examples.