Worth it… not worth it?
I did the deep learning specialization of the famous Andrew Ng. I hesitate a lot between doing it. There is a bunch of online courses in AI, deep learning, … Shouldn’t I just practice rather than studying, as I did for machine learning?
After deep thought, I ended up thinking that:
- Deep learning is a hot topic, I need to tick the box, somehow.
- I will hardly find a nice way to practice deep learning by myself. It’s pretty hard!
- The “certificate” will give me something on communication perspective, I guess. I am clearly an active learner and I need to “certificate” it.
After doing all the 5 courses:
It is a worthwile experience, clearly. The courses are very pedagogic, consistently structured. Andrew Ng’s videos are pleasant to watch. And the TPs are nicely done, they tried to avoid doing something boring, they are very interesting and open the doors of the standard framework such as tenserflow.
Now, they only open the door of this difficult-to-understand world. They give you like a nice flavour of what may be done but you see the (long) path to follow in order to really be an expert. This could be presumed.
This confirmed also my sensation that the deep learning neural network are powerful mathematical tools, clearly, but not to be used as a magic tool to “rule (them) all” the problems we want to solve. Indeed, you need many conditions to be able to profit from them. It is the case in image recognition, in NLP, with the huge amount of data and the type of these problems.
They may be adaptable from one situation to another, using some technics such as transfer learning. but it must be similar situation, with same inputs. One example is within image recognition, you might use a image recognition algorithm which recognizes well some low-level features, for a radiology diagnostic system where you have few images.
But I am not sure they may be used in an efficient way in many situations, as I have heard here, there and everywhere some time ago. It worth mentioning the following extract from this wikipedia article I had read when working on the meli data challenge and looking for the best algorithm to use:
While deep learning has been applied to many different scenarios: context-aware, sequence-aware, social tagging etc. its real effectiveness when used in a simple collaborative recommendation scenario has been put into question.