Hey, thanks for the comments!
GANs have definitely been achieving some unbelievable results and will for sure continue to be one of the most promising areas in ML in 2018. In addition, I personally think that we'll also see a lot of great work this year in: unsupervised/self-supervised learning, exploration, attention, and evolutionary strategies.
It's definitely not too late at all to move into ML! One of the best things about ML is the (relatively) low barrier to entry. In most other fields, in order to push the bleeding edge, you need to study decades of theory in order to reach the boundary of what's been explored. In ML, it seems that you can reach that point much faster, thanks to the wealth of accessible online materials and immense levels of open research communication.
So... where to start? I personally don't have experience with Udacity's nanodegree programs, but I've looked at the syllabus before and it seems reasonable. What I've seen a lot of people do is Andrew Ng's famous Coursera class, which is an online version of Stanford's introductory ML course.
Personally, I would do something like:
- Make sure I have a strong grasp of the fundamental prerequisites [Linear algebra, multivariable calculus, some mathematical maturity, Python]. If you are missing any of these, a few weeks diving deep would probably be good enough to pick up whatever else you need as you go.
- Take a course or two to get a structured overview of important topics in machine learning and deep learning. In addition the the Udacity program, I would also consider the Coursera class I mentioned above and the Fast.ai course.
- Consider a textbook or two. This is a classic.
- Projects projects projects! The best way to learn (and show others what you know) is to work on ML projects. Whether you choose to contribute to open source projects or work on your own, I think having a strong portfolio is one of the most important things to do as a CS/ML outsider trying to break in. These will speak far more than course certificates, in my opinion.
Regardless of what path you choose though... it's definitely not too late! ML is definitely here to stay. Calling it Software 2.0 might be slight hyperbole, but I have no doubt that ML systems (and people who work with them!) will be very crucial to how our society works in the next decades.
And yes, Google Brain is amazing. I love going to work every day with such smart colleagues and working on problems I care about.
Generative models for architectural design is definitely a very interesting problem! I'm not aware of any work that approaches anywhere near what you are proposing. I think a lot of the challenges stem from the very very complex dynamics that go into ensuring a building is structurally safe; while we may be far away from an end-to-end system that can output detailed, architecturally meaningful blueprints that are safe and useful, I think it is very plausible that deep learning systems can at least automate some mundane parts of the architectural design process. Which could very much drive prices down 10x in the future :)
Hope that was helpful! Happy to answer any other questions.
Seriously. Thank you so much. All of this info is unbelievably helpful and exactly the kind of direction I've been looking for. I started khan academy courses for multivariable calc and linear algebra for a refresher last week and this outline will help me immensely with my next steps. I just heard about Fast.ai in this panel conversation between Vivienne Ming, Jeremy Howard, and Peter Diamandis yesterday, and am excited to take the course. Dr. Ming's work has been super inspiring for me.
Thanks again for taking the time to respond so thoroughly! I will definitely take you up on that as questions come up through my studies :)
-Richard