Experience and Overview
This article is about my experiences while I was beginning to pursue a career in Data Science after graduation. At the end of every section you’ll find links to corresponding courses.
0. The Beginning
It was the final year of my college. I was determined to start my career in Haskell (a pure functional programming language) and become a type theorist in the distant future by pursuing a PhD.
1. Machine Learning — Andrew Ng
Since everything in my career was planned, I said to myself, “Let’s just take the course on Machine Learning by Andrew Ng and start a new career in Data Science”. And I continued being a proud member of the Magpie Developers Group. Here’s an article published in 2008 about it. And it still holds true.
It is a 12 week long course. The course is often called a bird’s eye view of Machine Learning (ML). The exercises are in octave. The course primarily covers three topics,
- Supervised Learning
- Unsupervised Learning
- Best Practices
Both the depth and the breadth of content is remarkable. Andrew is an inspiring teacher. One of his quotes got stuck in my mind.
Artificial Intelligence is the new Electricity. — Andrew Ng
I am very passionate about automation. Electricity brought the first wave and revolutionized everything. And I believe that Artificial Intelligence (AI) will bring the next. After graduating from my college, I had several options. Startup, Job, Masters, or PhD in the field of Data Science. I chose to stay at home and study Data Science full-time.
2. Reinforcement Learning — David Silver
While exploring the field of Data Science, I took several projects. I could not finish some of them. I tried using both Supervised and Unsupervised Learning. Neither was fitting in. That lead me to the third type of ML. Reinforcement Learning (RL). Andrew’s course did not cover that. I found a course on Reinforcement Learning by David Silver. It was accompanied by the book, Reinforcement Learning: An Introduction by Richard Sutton. Even though I worked on it full-time, I encountered enough problems to take more than a month to complete both the book and the lectures. I watched the lectures twice to understand RL completely.
There are 10 video lectures. The book has 17 chapters. A lot of research done on this subject seems to be documented in form of research papers only (see OpenAI). If you dig into this subject, do publish a few articles. Ping me on twitter @dhruvin2910 if you do. OpenAI’s gym, baselines and universe are great places to start, in that order.
RL in my opinion is closer to how we learn. By trial and error and by experience. RL algorithms can use function approximation, which makes them consumers of Supervised Learning. Whereas RL environments can use dimension reduction, which makes them consumer of Unsupervised Learning. Once the RL becomes mainstream, the programs will be godlike.
3. Deep Learning (Learning Path) — Cognitive Class
Courses mentioned above helped me gain theoretical knowledge. I have been programming the exercises in python, the usual scipy stack. It was time for implementing Neural Networks (NN) using tensorflow.
- Deep Learning Fundamentals
- Deep Learning with TensorFlow
- Accelerating Deep Learning with GPU (in beta at the time of writing)
The first course is a good resource for learning various types of neural networks (used in both supervised and unsupervised learning). It gives an overview of various NN frameworks in python as well. The second one dives deep into tensorflow explaining you how it’s done. The third one focuses on performance side of the NNs by explaining how you can leverage GPUs and Cloud (for both parallel and distributed execution).
After finishing the path, I realized that ML on cloud is going to be the future. Many cloud providers started providing ML as a service a long time ago. They offer pre-trained models (for predictions) for some of the popular tasks like speech recognition, translation, and object recognition. After the first stable release of tensorflow, it has become de facto framework for DL. It is a low level framework. Google announced keras as the official front-end for tensorflow. It is a higher level API and is widely used for rapid prototyping.
4. Deep Learning Specialization — Andrew Ng
The ML course by Andrew Ng was launched in 2013. A lot of new algorithms and best practices emerged since then. Machine Learning started becoming mainstream and Deep Learning was catching up. There was a need of a DL course (which would serve the purpose of the ML course by Andrew). Andrew and his team at deeplearning.ai launched Deep Learning Specialization in 2017. It uses python and jupyter notebooks for the exercises. It consists of 5 courses.
- Neural Networks and Deep Learning
- Improving Deep Learning
- Structuring Machine Learning Projects
- Convolutional Neural Networks
- Sequence Models
The first course focuses on the fundamentals of NN. It explains everything from shallow and deep neural networks to forward and back propagations using numpy. The second course focuses on hyperparameter tuning. It covers regularization, dropout, gradients, optimizers, and the list goes on. The third course is about project structure, error analysis and various practical topics not covered in the courses above.
The fourth course explains the ins and outs of Convolutional Neural Network (CNN). Its range spans from computer vision to various case studies and renowned research papers, which I found amazing. I waited a month for the last course to be released and let me tell you, it was worth the wait. It covers Recurrent Neural Network (RNN), word embeddings, speech recognition and various sequence models. The last two courses are the flagship of this specialization.
Topics mentioned in this section are just the tip of the iceberg. The courses are free to access. But this specialization is worth paying for.
5. Data Scientist Track — DataCamp
While waiting for the last course in Deep Learning Specialization, I became aware of DataCamp’s new year offer. I saw a few free courses and the projects and was impressed. Even though I had been using scipy stack for a long time, I had to google every time I got stuck somewhere. DataCamp’s python tracks seemed to cover from the foundations to best practices. They also offer courses in R. It’s a goldmine to me. So I subscribed. I could finish all 4 skill tracks and all 3 career tracks (consisting of 20 courses) within 15 days. #BingeLearning.
A track is just a list of courses. There are 4 skill tracks (in python):
- Python Programming (4 courses)
- Importing & Cleaning Data (4 courses)
- Data Manipulation (4 courses)
- Machine Learning (4 courses)
Related courses are grouped as skill tracks. The first skill track focuses on the basics of the python language, numpy, pandas, and matplotlib. It covers syntax, semantics, functions, erros, scoping, arrays, data-frames, and plotting. The second skill track is about getting data from CSV, Excel, Web (scraping), APIs (JSON), SQL (databases), and cleaning data using pandas.
The third skill track covers pandas foundations and sqlalchemy. The fourth skill track covers supervised and unsupervised learning using scikit-learn as well as deep learning using keras. It has a course on ML from experts, which breaks the myth of only complicated models winning ML competitions.
There are 3 career tracks:
- Python Programmer (10 courses)
- Data Analyst (10 from Python Programmer + 3 = 13 courses)
- Data Scientist (13 from Data Analyst + 7 = 20 courses)
A career track is list of courses to complete to be able to start a career in respective field. I started from Python Programmer and worked my way to Data Scientist. Shorter goals help, and I’d recommend you the same.
I liked DataCamp’s exercises. The checker is pedantic sometimes but it never got in my way. You get a python environment with all dependencies installed to do the exercises and check them online. DataCamp’s Data Scientist track ends where Andrew’s Deep Learning Specialization starts with some overlap.
6. Data Scientist — DataQuest
I learned a lot while completing the above courses and specializations. Currently I am learning from DataQuest, a platform similar to DataCamp. DataQuest seems to offer a lot more than just the Data Science specific content.
This is my first article. Before writing it I thought documenting my journey won’t be that helpful to others. I shared a draft with my friends and the responses were positive. Thanks Varun Barad and Dixita Ganatra.
Divyesh Peshavaria mentioned that I should include details of the projects as well. I am thinking to publish an article for each project in near future. Maybe a series.
It’s been 8 months since I graduated from college. I’m planning to find a job and continue learning afterwards. Now is the time to put my skills on test.
Source: Deep Learning on Medium