Deep Learning has taken the world by storm and the juggernaut has kept rolling since early 2017.
As far as the core methodology goes, neural networks have been around since decades and convolutional neural networks and recurrent neural networks have been around since 15 odd years.
What has changed suddenly you ask? GPUs and breakthroughs in automated systems like self driving cars. As Andrew Ng himself says, “I think the other reasons the term deep learning has taken off is just branding. These things are just neural networks with more hidden layers, but the phrase deep learning is just a great brand, it’s just so deep.”
Developers all around the globe are heavily motivated seeing the innovation DL is driving in each and every sector. Every company is either pitching themselves as providing Artificial Intelligent solutions or a blockchain company. Quite frankly, these 2 things have become the Guitars and DSLRs which college going students use to impress people but very few take it to the next level.
So how do we get our skin in the game?
I will try to analyse some of the most popular and widely accepted online courses for deep learning as I have almost completed them all (I am yet to finish Udacity’s NanoDegree’s RL part, which itself is a discourse on its own) and break them down in terms of the value they add versus the cost it takes out of you — time and money.
We will discuss about 4 courses in totality:
Lets begin with a short intro to all the courses:
Deeplearning.ai, Andrew Ng’s boat which leads us towards autonomous salvation is split over 5 courses covering NN, CNN, RNN and dedicated courses to hyperparameter tuning and ml project structuring. The courses contains 3–4 weeks respectively(except course 2 and 3) which contain as many quizzes and programming assignments in python hosted on coursera’s jupyter hub.
DeepLearning A-Z on Udemy provided by folks at superdatascience.com is a 23 hour rigorous course which splits the course into 3 parts : Supervised learning(ANN,CNN,RNN), Unsupervised Learning( SOM, Boltzmann Machines, Autoencoders) and some basic ML referesher and data preprocssing techniques and templates.
Applied AI with Deep Learning by IBM is focused more towards the IoT/ production level design of deep learning systems with introduction to various frameworks like deeplearning4j, systemml and tools like Apache Spark and IBM Bluemix platform which makes exploring deep learning on real time data. Its a 4 week course which covers a lot in lost less time.
Udacity’s Deep Learning Nanodegree covers everything from NN to GANS and previously was marketed as a foundations course but they added Reinforcement Learning and dropped the former name. Contains rich video content and hands on mini segments which are presented in the form of self paced jupyter notebooks. Assignments are reviewed and mentors are also assigned for a personalised touch.
So, after a brief introduction to all the nominees for the Best Course on Deep Learning, the Oscar goes too…
Hold on, lets take this to the jury room.
From a learner’s perspective, I came up with an evaluate-tionary framework of auto M.A.T.I.C.S (Money Assignments Time Intuition Content Support)
-Registered in the state of Beliefotopia. All hail our saviour Geoff Hinton and his scribe Yann Lecun.
Good things come to those who wait and the metric goes in the increasing order of relevance as well, so stick around!
Lets discuss all these points with the perspective of students(S) and working professionals(P)
S: — Coursera is known for the financial aid system it has in place to help students not only get its content but also get certified for the course for full or partial payments. Udemy’s courses are comparatively very cheap (Less than $20) and the content might be available on the internet if you look closely. Udacity is probably the only one which provides free alternative to its nanodegree course content but then it charges $300–$500 dollars for its nanodegrees for the personalized code reviews and mentors.
P: — If your company is paying for your courses as part of L&D policy, I’d say do them all! If its your own pocket through which you are looking to pay then I’d suggest having Coursera’s $49 per month full access to their catalogue. Udemy is just a day’s salary for you and Udacity probably worth 2–3 weeks.
Verdict:- If you don’t want certificates but only knowledge then audit coursera courses, skip udacity nanodegree and be a pirate, harry! Else, shell out some bucks for the course creators’s hardwork and be a responsible student.
Udacity beats everyone in terms of assignment quality and code reviews. Andrew Ng has provided some cool Ipython notebooks but I must say, the assignments are not at all intuitive. Udemy course provides some very hands on intuitive level assignments/explanation where the tutor explains each and every line of the code which does help if your theory is on point. IBM’s course provides weird yet interesting assignments as they try to bring IBM Watson platform and big data aspect into real time production level assignments but are poorly explained/structured with weird UX around IBM platform.
Verdict:- If you like someone explaining all the code, Udemy assignments are very good. You want your code to be reviewed/mentorship — Udacity is bang on. For everything else, there is Coursera.
S — Students are at an advantage for they can easily pace out the courses over time. Coursera 4 week courses if done properly hardly take a week if you are clear with your math fundamentals. Udacity has much more to offer in terms of mini assignments and longer syllabus which justifies its ask of 8–10 hours per week. IBM course is very swift but it covers very less things. Udemy course is around 23 hours runtime which clearly takes atleast 2 weeks to finish.
P — All the courses are also kept keeping professionals in mind. Udacity has some deadline based projects so you kinda have to keep up. But burning the midnight oil everyday is better than chugging through the whole content over the weekend for you will forget it the next week.
Verdict:- Coursera is value for time, Udemy is self paced but might be frightening at times. Udacity is not for the faint-hearted if you are not well versed with its prerequisites of math and python
Andrew Ng covers the underlying math in a much much better way than all of the tutors but at times it proves to be counter intuitive if you like coding and hacking your way through it.
Udemy course covers both code logic and tech intuition in a nice and precise way but visually is unappealing.They also provide links to research papers to read which can be used as supplementary material to strengthen your knowledge on a particular topic.
Udacity content is very intuitive though they provide some write ups and math below their videos, some math intensive sessions (RNN) surely are snoozeworthy.( Important, yet snoozyy)
IBM course gives you an idea about IoT and production level deep learning with its very basic and intuitive videos which are good if you are looking to quickly try and build something without going into the grittiness of deep learning.
Verdict:- I personally found Udacity content to have cleared a lot of mydoubts, but then Udemy clears your coding intuition and Andrew Ng’s lectures though less intuitive but are necessary if you re looking to understand how everything works. You can choose to go for Udemy if your aim is to prototype rapidly and learn intuitively only.
Andrew Ng’s content is very standard and it feels as if you are sitting in a classroom and your professor is breaking down each and every difficult concept into simpler terms.
Udacity’s content is graphically attractive and doesn’t make you sleep( RL and RNN will) coupled with special content by Andrew Trask, Ian Goodfellow and Sebastian Thrun keeps things interesting from time to time.
SuperDataScience team has come up with their hit A-Z series in many verticals of AI/ML/DL and though their slides are not a visual treat, their content is precise and inclined more towards the practical aspect. They provide some very cool templates which are heavily reusable if you plan to experiment.
IBM content is a bit hasty I would say, the varied pace at which different instructors go and the accent of the main tutor is sometimes very difficult to comprehend without subtitles.
Verdict:- Udacity — Yayyyy, Udemy — Yayyy, DeepLearning.ai — Yayy, IBM — Yay
I’d say Udacity’s 60% fees might be justified for the support they provide. Slack channels, weekly office hours, live webinars with top scholars, mentors to help you out with code and logical errors and portfolio building skills make taking a mooc a very wholesome experience.
Coursera and Udemy have support forums where you can discuss with your fellow classmates or reach out to TAs but then it depends on how much grasping power you have. Better the support, Better the understanding.I personally got frustrated with the IBM course as the assignments are so rigid I almost gave it up at one point just because I was copy pasting wrong json file.
Verdict:- Udacity takes the cake and justifies its price in this segment
Its one thing to download tutorial videos and other thing to take an actual course. I have had like 30–40 gb worth of tutorials ranging from Game Theory to How To Make Delicious Pastas and I still eat instant pasta packets. The point being if you are commiting yourself to a course and are serious about entering this field or are just wanting to have an overview of the fuss in the market, I have some yoda wisdom to pass on.
- Chronological Order : Deeplearning.ai, Udemy, IBM, Udacity
- Machete Order: Udemy, IBM and then watch the world burn
- The Chosen One: Udacity DLND
- Original Trilogy — Andrew Ng’s Machine Learning-> Khan Academy Math ->DeepLearning.ai