Just finished it - would recommend to people who want a bottom up approach for learning deep learning algorithms. It definitely complements Jeremy’s top down approach and helped me to understand couple of “why” questions.
NB - There was supposed to be a miscellaneous thread where I could post this, but couldn’t find it - sorry for that.
Especially creating Conv nets from scratch in Numpy…
Totally subscribe to this point of view.
I loved these courses but have been waiting forever for the last part on Sequence models. I took those courses first before coming to fast.ai. To me, I like that progression as Andrew Ng does a great job on concepts and basics but implementation of real code is missing whereas fast.ai is amazing at showing how deep learning experts actually do things.
I’m waiting for the Sequence models course too ! I agree with the implementation of real code, even though, from what I see so far (only on week 2) I’d love to have a detailed of fast.ai code, how it calls pytorch etc.
Well you can ask your question out here - https://discuss.pytorch.org/ . And for detailed explanation, you can ask people in this forum as well.
I wonder how many want Jeremy To Teach Python??
I am first one who will be glad if Jeremy creates a lecture or two on Python(tricks, OOPs etc…)
Especially the way he taught the OOPs concepts,(lec 6,7 ML)
I can bet that no book I know of did it that way…
You are Awesome…
May be tag Jeremy here - he might respond to your request.
I’m here It’s an interesting suggestion… I’m really glad you liked the mini OOP thingy
Can you share link to that courses? Thanks!
All the lessons (4 courses) are also available on youtube.
Deep Learning .AI
@mmr Could you give some examples of “why” question you mentioned in the original post? Thanks.
Well there were couple of things explained in much granular format using numpy -
e.g - As someone already mentioned inner working of convolutional network in numpy.
I was interested in knowing how some of the weight update algorithm works - RMSProp , Adam, weight update using momentum.My favorite was actually face detection and face recognition using conv net - it was an application of one shot learning.
But I still have to maintain that take the best of both of the two worlds - lessons from fast.ai Part 2 - as well as deeplearning ai - you are good to go in deep learning.(…thats my subjective view)
What I found that , if you have done the courses here, you should be good to read scientific papers and implement then yourself without much effort. There was not a ton of value that I got from Coursera. I found that the assignments were too trivial and to much was set up for you. There were one or two things that I liked, like some things in pure numpy. If you do want completeness you can read the deeplearningbook and watch the lectures for that. It’s covered on a chapter basis.
But scientific papers go into mathematics always…?
How to overcome that?
We can now implement a model as shown in the architecture but what they add to it makes it difficult to complete the paper implementation…
Things whizz pass my head just like a Bullet..
Most of the math in papers is just difficult notation but logical easy to grasp.
I still dont understand a paper 100% the first time i read it, but after a few times it gets fairly easy to grasp
Couldn’t agree more. The two courses work great together specially for beginners.