Learn BERT - essential NLP algorithm by Google (Udemy.com)

If you're considering purchasing this course, don't. The course content is very outdated and when you ask questions in the Q&A, all you get is an automated bot that tells you they can't answer. If it wasn't for other students' contributions, I would still be stuck on the first application. Thus, I spent more time debugging issues related to the dependencies being old than actually being able to understand what I was doing; and due to computing limitations in colab, I was also not able to get any results in two out of the three BERT applications built.

I've found AMAZING the first theoretical part of the course so we can understand recent development history on deep learning, how BERT works and what's new about it (pre-trained model, bidirectional context, attention masks). On the second part, more practical, you follow the construction of BERT layer from scratch for a SQuAD problem, which is pretty good, but for me it was a bit overwhelming since at work I'm just using Transformers Trainer API or building on top of PyTorch or Tensorflow specifically on BertForSequenceClassifier to solve our business problem on e-commerce classification of products and queries.

The course is good and one can start using BERT afterwards., provided that the student practiced with notebooks. On the negative side: * I find the explanations not sufficient and little visual. Additional slides would be appreciated. I am a person who better understand things by seeing as opposed to hearing. * no live coding please! this is a waste of time. It would be more useful if all code is already in the notebook and the teacher goes through the code and explains it. This would allow, among other things, to explain more details in the time saved by not typing. * I would appreciate if the teacher runs the notebook cells and shows the outputs more often than he does. Ideally a dedicated demo for each function. Other than that, thanks for the course! It was useful.

I just watched the lesson explaining word embeddings. I have read many explanations, often using the same or very similar graphics, but this is by far the best explanation I have ever seen. It breaks the whole thing down into terms and relationships that are clear and readily understandable. It makes me think that some of those others classes didn't understand embeddings at all, and so just repeated the technical language of the original authors to hide their own ignorance. Bravo!

As an NLP practitioner, I found this course fascinating: the application of BERT first as an word-sequence embedder to a CNN and then as a QA context labeller was really interesting and useful. The course was clear and well-paced. I also liked the attention to both intuition and practical parts. However, I would say that the end of the QA task was a little disappointing as it: (a) covered a lot of the preparatory material for handling SQuAD files in a very rushed manner, not giving much instruction as to what was going on. (b) Once the evaluation of the QA application was done, there was no investigation of individual data instances to see how the QA performed in practice and what mistakes it made. The instructor alluded to additional material that might cover this. But it has not appeared.

I like this course! I appreciate the level of detail that instructor shared to the coding section and his insight and experience sharing. Sometimes a couple sentence of explanation save me quite some time of web search. I like the project example with actual hold-hand coding. I admit initially it's harder to follow since I am new to this area, but it gets easier as the 3 projects are related to each other and can carry on to the next project. I have a solid learning feeling. The only thing is if the transcript could be real but not auto-generated would help me to understand some word or terminology easier. This is one of a few courses I've actually completed end to end.

The material was well explained and I recommend this class. The reason I didn't give it 5 start is that the fine tuning section was done on a very specific task that is probably not helpful to the majority of the people. It would be much more useful to continue demonstrating everything on the classification task, including the custom training which was good to see. Choosing this very unique and complex task didn't contribute anything to understanding BERT better.

Yes wanted to learn BERT for quite a bit and till now the intuition developed works well. Still, a long way to go but would like to explore further. Update: Replies have not been on time, have been trying to get questions answered for 3 days now due to which I am stuck, and cannot move forward in the course.

I find it rather ambiguous when the instructor mention Shape and dimensions of the layers in the network. I know he did mentioned it in the first part but still, the learner has to go back and find the diagram and it really affects the flow. So I would suggest while constructing the network, he could add a diagram to illustrate the dimension of each layer.

I was hoping to learn everything about using BERT, but I haven't. The most important things (how to prepare data and what is returned by BERT) took just a few minutes without ANY visuals. Don't show your typing, show a well-commented code and tell why did you write this and that. And don't use SQuAD, your students will have their own tasks!