Advice for What Area to Pursue in Masters
Hey TeamBlind, was wondering if I could have some advice. I graduated from Berkeley this spring, 2019, and interned at Amazon and am currently re-interning at Amazon (yeah I know, but I really liked it, was on AFT and it seemed chill other than the Prime Day Fiasco and on-call). I was going to graduate early so I decided to doing the 5th year masters which is a research based program where you conduct research, take grad courses, and publish a thesis/publications :https://www2.eecs.berkeley.edu/Pubs/Theses/Years/2019.html#5th%20Year%20M.S.
I was wondering if you guys had advise on what area to pursue. I was interested in machine learning with a focus on computer vision and robotics before, and did undergrad research in this area. However, after interning at Amazon, I was impressed with the infrastructure and complexity of everything, it was eye-opening to peek behind the certain and see how much work goes into it. I took an operating systems and database class and am currently interning on AWS. I love keep seeing how the stuff I learned in those classes keep coming up, and in real life practice! I love how we're pushing the boundaries and have real world impact with our work.
Personally, I think I would like to go more into the software/systems side which Berkeley has really good research in too. But there is a side of me that's hesitant - having the opportunity to do more research in deep learning and computer vision, it like turning down a ride on a rocket ship that I know many people would kill for. I did at one point and worked hard to get that opportunity. Berkeley has become a world leader in computer vision and robotics, it is such an amazing opportunity to be able to work with brilliant experts in the field pushing the boundary of what's possible.
So that's the dilemma. Two great options that I can't choose. What do you guys think?
* Computer Systems: https://ucbrise.github.io/cs262a-spring2018/
* Machine Learning Systems: https://ucbrise.github.io/cs294-ai-sys-sp19/
* Really get to dive deep and become an expert in in the literature and theory of building complex systems that scale and are fault tolerant
* Strong practical experience and is applicable to any technology company and has high impact
* Probably has higher demand and safer in a recession as ML can be surplus as opposed to keeping systems up that generate money
* Can be really difficult and tedious
* Not as hyped as Deep Learning and might grow stagnant
* Miss out on the opportunity of working and learning from some of the best in Deep Learning
* Deep Learning: https://bcourses.berkeley.edu/courses/1478831
* Machine Learning: https://people.eecs.berkeley.edu/~jrs/189/
* Deep Reinforcement Learning: http://rail.eecs.berkeley.edu/deeprlcourse/
* Computer Vision: https://inst.eecs.berkeley.edu/~cs280/sp18/
* Human-Robotic Interaction: https://people.eecs.berkeley.edu/~anca/AHRI.html
* Deep Unsupervised Learning: https://sites.google.com/view/berkeley-cs294-158-sp19/home
* Opportunity of working and learning from some of the best
* Disruptive technology that can impact our way of life and many industries in the future
* Get to be on the frontier of what's possible and maybe push the boundary a little further
* While I'm decent at math, this is probably a weakness I'll have doing research work
* Actual research/machine learning engineer positions are scarce
* Impact is more uncertain, projects can be big successes or failures
* Has a lot of overhype
I think my personal reason for learning towards software/systems was seeing the disconnect between people in college and industry. I've seen many other CS students overhyped on ML whereas when I went to talks and talked to people working on ML at Amazon, the actual model building and theory pushing the boundary was a small part (Google might be different for example). There's a lot of other work, data collection and wrangling, building pipelines, deployment and user studies, etc. Even the model building often is just using libraries. Academia focuses so much on new SOTA model that is 2% more accurate but 50x more computationally expensive when in industry simple models work well enough.
And of course, since this is the blind circlejerk, which has higher TC? (Don't care as much both are good and there's probably no certain way to know).