"machine learning in computer vision" VS "machine learning in computer architecture"

Feb 14 12 Comments

Hello I am a first year PHD student in the field of machine learning. I am choosing the topic between "machine learning in computer vision" and "machine learning in computer architecture". I am wondering if computer vision has more application and it can lead to a job easier than the field of computer architecture after graduation in 3 or 4 years? It seems that computer architecture is essentially ASIC/FPGA design and it belongs to hardware category, and it seems there are not much hardware position in the market compared to computer vision, is it correct? And also is it true that hardware engineer/researcher (even in NVIDIA and APPLE) are not compensated at the same level as the SWE or MLE with the same years of work/research experience?

If anyone happen to know about these two fields, can you please shed some light on this? Because choosing a topic is a big commitment for a PHD, it is very hard to change the topic again after a few years in. Thanks

comments

Want to comment? LOG IN or SIGN UP
TOP 12 Comments
  • Intel bannon
    If you are starting your PhD , don’t go just by the hot field. You will be in the program 4-5 years and the world will change a lot by then.

    I know people who did their PhD in VLSI because it was hot but they aren’t making as much money as the CV guys. Guess what CV was looked down upon at the time and people had trouble finding jobs. So, don’t try to plan is to the finest details and try to do really quality research with a rockstar prof.
    Feb 14 3
    • Agree the system always overshoots before it stabilizes
      Feb 14
    • OP
      Hello bannon, I wondering is it that VLSI was just as hot a few years ago like CV and ML today?
      Feb 14
    • Intel bannon
      It was not. But it was the hottest thing around. I think ML as the area is good for a long time. Inside that, I think you will make a mistake in overthinking. If you want to work in industry, it is very unlikely that you will get to work on exactly what you will do in your research. If you have the broad skill set you will be a hot commodity. So I would not put my chips on the finest of subareas and hope to make hay out of it. You don’t realize now, but PhD is a big commitment and you are unlikely to get out in less than 5 years. So I will repeat myself, find the brightest and most ambitious prof in your department who works on ML and just go with it. Good luck!
      Feb 14
  • Nvidia tewqer
    I think it’s very important to chose what you enjoy for your PhD topic. You will do well in either of them. But there might be more marked demand for the software , algorithm and application aspects of ML than computer architecture. Both are great fields though.
    Feb 14 0
  • Microsoft Girlfriend
    CV just means you’re applying the ML model to primarily images, signals, (stuff the computer can “see”) and architecture is pretty much everything else. The foundation is all the same. When you look for jobs they’re going to ask you the basics, perceptrons, linear SVM, other regression methods, backprop, then they’ll have you leetcode, and once you pass you get put on a team. If you want to work for Nvidia you need to be great at c++ and understand kernels etc. if you want to work for literally any other tech company in ML, you need to know python and all the popular libraries i.e tensorflow, scikit, torch, etc
    Feb 14 0
  • Amazon lonebhezos
    CV is a solved problem already, and it’s saturated.
    Feb 14 2
    • OP
      Hello lonebhezos, I am wondering if you are sure that CV is saturated because it seems there a still quite a lot of work about algorithm optimization that can be done? It seems that the training time for CV ML model may take days even weeks?
      Feb 14
    • Microsoft vMEa72jdje
      Hahaha. Have you ever heard about CVPR?
      Feb 15
  • Dell atxdell
    Even if you believe CV isn't yet solved it will be in 4 years when you finish. Architecture is about researching methods to accelerate such as fp32 vs fp64, distributed model training etc and how that affects ml performance, accuracy etc. The PhD will be responsible for turning those in to.designs such as tensor cores and nvlink.

    Tons of company's hiring these roles as the cost to produce unique semiconductor designs drops companies are looking for PhDs to help them develop novel computation models to leverage in silicone to have a competitive advantage.
    Feb 16 1
    • Salesforce AkOj58
      AI/ML is not going to go away for the next 10 years. Very little has been done yet
      Feb 17
  • Salesforce AkOj58
    Cv and ML. Forget about anything related to hardware.
    Feb 14 0