إلى طلابنا في حاسبات ومعلومات - تعليم الآلة - 3 - [Machine Learning - Learning roadmap] смотреть онлайн

My ML Live Course: https://docs.google.com/presentation/d/1dcvabQyk11ZaP746E89UyVwdHmuESLYUL2Kk4cfBAAk/edit

Roadmap for learning Machine Learning (More focus on DL):



- General Notes
-- Roadmaps nature
-- Recall DL vs Classical
-- Paid courses & certificates - financial aid system
-- MOOCs assignments
-- Can't understand something?

- When to start learning ML

- Roadmap as following
- Andrew Ng Coursera Course
-- https://www.coursera.org/learn/machine-learning
--- https://www.youtube.com/playlist?list=PLLssT5z_DsK-h9vYZkQkYNWcItqhlRJLN
--- https://medium.com/analytics-vidhya/python-implementation-of-andrew-ngs-machine-learning-course-part-1-6b8dd1c73d80

- Kaggle
-- Browse scikit-learn library - use their NN tool (and others if want)
-- https://blogs.mathworks.com/loren/2015/06/18/getting-started-with-kaggle-data-science-competitions/
-- Kaggle: titanic (classification) and Boston housing prices competitions (regression).
--- https://www.kaggle.com/c/titanic
--- https://www.kaggle.com/c/house-prices-advanced-regression-techniques

- Go deeper in Neural Networks
-- Why NN?
-- Deeper understanding
--- Backpropagation
--- You may listen to 1 or more lecture of these:
--- https://www.youtube.com/watch?v=Ih5Mr93E-2c
--- https://www.youtube.com/watch?v=uXt8qF2Zzfo
--- https://www.youtube.com/watch?v=d14TUNcbn1k
--- On paper: Compute/Trace an example: https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/
-- Reimplement NN from scratch
--- Multi-class classification (softmax)
--- Multi-label classification (sigmoid)
--- Regression
--- Try AE with your network
-- Learn dropout and implement it
-- FYI: NN/ML Arabic playlist: https://www.youtube.com/playlist?list=PLQkyODvJ8ywsLydDYORIlJxV9KarhXp9O
-- Maybe read NN chapter from a book

- Evaluation metrics
-- https://towardsdatascience.com/metrics-to-evaluate-your-machine-learning-algorithm-f10ba6e38234
-- Later: https://medium.com/@jonathan_hui/map-mean-average-precision-for-object-detection-45c121a31173

- AFTER you finish the above, do the following time split strategy
- Next 25% of time for classical/projects - 75% for deep learning

- 25% of your time for classical ML
-- kaggle projects / Kaggle blogs
-- https://www.coursera.org/specializations/aml
-- Google ML doc: https://github.com/hktxt/bookshelf/blob/master/Machine%20Leraning/Rules%20of%20Machine%20Learning%20Best%20Practices%20for%20ML%20Engineering.pdf
-- Feature Engineering for Machine Learning: Principles and Techniques for Data Scientists

- 75% of your time for deep learning
-- https://www.coursera.org/specializations/deep-learning
--- https://www.youtube.com/channel/UCcIXc5mJsHVYTZR1maL5l9w/playlists
-- [FYI] https://classroom.udacity.com/courses/ud730
-- Frameworks (TF/Pytorch) - study while learning
-- Either learn 2D vision or nlp (later after work learn 2nd)
--- Vision key problems: classification, detection, segmentation, pose estimation, object traction, action detection/recognition
--- stanford courses
---- Vision: http://cs231n.stanford.edu/syllabus.html
---- NLP: http://web.stanford.edu/class/cs224n/

- More on the road
-- ML book, e.g. Bishop
-- Understand more models, fields, more experience
-- Field follow up: E.g. in Vision (CVPR, ICCV, ECCV)
-- DL Experience: Paper-to-code skill
-- Vision elements (later): https://classroom.udacity.com/courses/ud810
-- Abstracting skills

- Other ML areas (later)
-- Generative models
--- PGM, EM (GMM), VI, MCMC
--- DL: VAE, GAN - Study GAN)
---- https://medium.com/@jonathan_hui/gan-some-cool-applications-of-gans-4c9ecca35900
---- https://github.com/nashory/gans-awesome-applications
-- Reinforcement Learning

--------
Updates/Additions
-----
- What if you are also interested in competitive programming?
-- ML is time-consuming to be good. In this case, divide your vacation to 50% competitive and 50% for ML