I agree with your advise of joining a smart city company but before doing so, what non-CS skills should I possess? What I wish to do is to accelerate the development autonomous vehicles from Level 3 to Level 5. One path or in my opinion the only optimal path towards the development of Level 5 autonomous vehicles is by constructing smart cities and fully integrating autonomous vehicles with smart cities. Without this approach, I believe that the development of Level 5 autonomous vehicles will be almost impossible. Furthermore, I can confidently say that China will be the first country to build Level 5 autonomous vehicles.
Also, you are talking about deep learning and how it is being widely used today. I understand deep learning and I am working on a project that is extensively using deep learning. However, what I want to do in the future is to build fully autonomous systems. The deep learning that we mostly see is nothing but a glorified pattern recognizer, brute-forcing its way through datasets to produce results, much like an overenthusiastic intern trying to solve every problem with a hammer because that's the only tool they know.
I would familiarise as much as possible with control engineering as far as possible.
I do have some books I could suggest, but they are very specialised/antiquated to engineering+maths (and really I learned lot more on the fly in practical domain since).... you will find better ones if you search w.r.t more computer science domain on it.
Its all about handling the degrees of freedom in the systems nested within each other....to apply proper feedback loops etc so they are say input-sensitive, output-dampened to be optimally self-governing as far as possible w.r.t the operational environment (and its imperfect perturbations or output patterns generated etc). Any and all science and engineering related to this that you feel will be important.
This will give you a leg up for your final year of CS I think and then career after.
You can try your hand at writing some routines of representative units regarding implementing these control forces within them and then between them.
i.e some small scale simulation of first an idealised autonomous car and then having some number of them in a circuit flow (representing a traffic flow etc) for them to communicate among each other and realise optimal patters for that flow....i.e the inside out stuff.
Then you introduce imperfections of the real world to test the responses and so on and iterate to improve....and grow the flow size capacity and complexity too as you can. Only much later stage will you get to think about how to extract more useful data from a larger simulation by "throwing" a large number of imperfections represented by something like a monte carlo sim....and see the larger systemic response to it in outside-in fashion.
The larger you are able to grow the flow, the larger the numbers and then you might gain insight then into where deep learning can be implemented to handle scaling automatically and so on (by sandboxing in smaller scale, and designing algorithms to optimally filter the sensitive factors that are most relevant for scaling in the larger control design loop).
This kind of thing will naturally induct what you need to read more on and you will gain insight into the best specific subject matter experts of note (that have worked on it so far) most relevant to you, i.e source, names compiling in a log....and which alleyways of further knowledge or practical implementation you sense most utility and interest in relative to say what you can bring to a company you research and later join etc.
This is just from my approach/experience....I am biased by my field and coming at it through that lens.....i.e thinking of the larger flow (that you want to study and produce results for quicker through say AI). To me for autonomous cars etc, essentially the control logic inside and between them is the same kind of thing as forces between gaseous molecules essentially.
LOL: )
