I have always been quite confident in my learning ability. I believe for any given topic, I would be able to understand the logic behind and apply those knowledge to solve (sadly, but true) exam problems within a short amount of time. From school to university, even until early this year.
This level of confident has decreased recently in a good way I think. It reminds me that there are topics in field even I am most familiar with and interested in require significant effort to understand not to mention master or even contribute.
The two courses I taken via Coursera from Sep are Neural Network & Probability Graphic Model. Most of my learning is to watch the lecture videos while I commuting between home and office. I didn’t put in enough effort to really read through the slides and research on further materials mentioned by the instructor. I found myself missing quite a number of knowledge point. Especially in NN, many incitation and techniques introduced are very alien to me. Not only I can’t understand why do they use those tricks to train a network, but the exact difference between techniques.
Quizes for these 2 courses are still quite manageable, even though I am not fully understand what the lecturer said. However, those programming assignments are really big headache to me. Sometimes I spend more than 10 hours on 1 assignment, and still cannot figure out the right way. Sometimes I have to rely on reverse engineering by guess the expected result via trial and error.
It is a bit disappointed when I read that the answers for those programming assignment will not be released even after the course ended. It would continue be a status of half-half for me, even I can score perfectly by bootstrapping the answers.
After a few more hours on the assignment last night, I decided to stop. I will keep the assignment in my “Someday: ToDo” list when I have the opportunities to really go through the material thoroughly again.
Back to the original purpose of this post, I am here to remind myself that in order to be an expert in an even narrowly defined field, I need to prepare myself mentally. Knowledge and skills in this world are so diverse and profound that so many smart people have contribute their lives. I should not expect a shortcut to lead me there.
EDIT: Turns out the NN staff point us the right direction with an excellent post in the forum who carefully explained the algorithm.
And it turns out that I was distracted by the ret=loss() function provided in the sample. My weight decay is totally off. I should have use what I learned from the Machine Learning by Andrew Ng on the regulation term.