ML_Project_Notes - RicoJia/notes GitHub Wiki

1. ReID: eventually "suggest" with confidence.
   1. make some recordings, even using bag labeller
   2. how to split the training and testing data? maybe cross validation? mismatches during learning, well... right now we start evaluating once there's at least one sample under that track id.
   3. Classifier:
      • nit:
        • when do mismatches happen?
          • misclassification (svm)
            • tyr ovr
          • camera view is faulty
            • (could there be a swap?)
            • camera view captures part of the figure
          • camera view is not complete and it happens to look similar
          • Dark region case (Filip and I look similar)

========================================================================

========================================================================

1. What do we want from machine learning?

   • What do you like:

     • Build a product that directly brings concrete value. But my product, doesn't have to be one single product??
       • Robotics:
         • Reinforcement Learning; Planning with robot
         • Computer vision: it's less of "issueing actions", but is an extremely important aspect. I think I'd like to keep growing in this aspect. I've worked on ReID (simple machine learning techniques), human tracking systems (particle filtering and various velocity models based).
       • Non-Robotics
         • Recommendation systems: I think this helps people explore things more to their likings, while helping a business grow.
         • LLM: A startup idea I saw was pretty cool: filling out immigration forms.
       • ZERO passion for optimizing ads - it doesn't create concrete values in my opinion.
     • Eventually, have my own company

   • I want to become a machine learning engineer. Because:

     • The world is moving towareds more AI, well machine learning based approaches. It's crucial to know conventional control theory, optimzation methods, but machine learning is where the spark is, I think
     • Robotics costs a lot: an AMR costs $4000 (like fetch base); Kinova arm: $25000; Nuc: 500... For machine learning, you can get started on a $300 GPU
     • In my current job, you spend a lot of time testing on hardware. I don't feel like I'm working on actual AI problems ()

   • What was the thing that pulled you into Robotics?

     • Build something that physically change people's lives. Also, I thought control engineering was really cool.

   • Do you still like it?

     • Yes. But building a robust system is costly. I need $$ for building my own robots, and my own company.
     • It's likely not profitable for a long period of time, either.
     • To make it robust could be very tedious. I feel like I'm out of touch with machine learning.

   • What was the thing that you DIDN'T choose Machine learning in the first place.

     • I was focused on understanding robotics theory, control systems, kinematics, programming. So I can spin up a robot. But now, ML is more integrated into robotics for more computational power.
     • Optimizing ads is lame ...
     • Optimizing Uber, google maps; chat GPT feels like a lot of people are working on them, already. It's pretty saturated.

   • What changed your mind?

     • Recommendation system could be interesting. It's very profitable. However,

   • Tell me about what you know about a machine learning engineer's life

     • 90% time of data cleansing, preparation, you don't really get to build models yourselves. You train the models, which could also be a tedious job.

   • What aspects of machine learning do you want to work on? The ones in robotics already?

• Intro to Machine Learning https://www.udacity.com/course/intro-to-machine-learning–ud120
• Machine Learning https://www.coursera.org/learn/machine-learning

========================================================================

========================================================================

• http://www.juyang.co/phd%E8%BD%AC%E8%A1%8C%E4%B9%8B%E8%B7%AF/
  • Must consider what you have; If you don't know NLP, but just how to run models, you don't necessarily see the key pain points.

========================================================================

========================================================================

• linear algebra, differential vector, calculus
• numpy & pandas: https://www.udacity.com/course/intro-to-data-analysis–ud170
• statistics: Intro to Statistics https://www.udacity.com/course/intro-to-statistics–st101 Intro to Descriptive Statistics https://www.udacity.com/course/intro-to-descriptive-statistics–ud827 Intro to Inferential Statistics https://www.udacity.com/course/intro-to-inferential-statistics–ud201 Time Series Forecasting https://www.udacity.com/course/time-series-forecasting–ud980

• Harjatin:

  • Robotics: dataset hard to find; Saycan:
  • DATA; Sim2real: data problem: randomize in simulation. Learn everything (friction coeffiction,
    • model based / model free RL wind policy.) Not working; not with start ups; So much money;
  • Tesla: machine learning, so they are able to do cameras
  • Visual servoing, button pressing: RL. How to get data.
  • Market robots: nobody is

• Software: glass ceiling. It's easier to do

  • Reward.
  • Computer vision; go stores: shutdown;
  • Things took off; LLM: afterthought; Has to happen in robotics. 10x cheaper;

• Next things in Robotics:

  • Data2Real; Machine Learning Based. Train Everything. RL - simulation will

    • Model free RL.
    • Sergey Levine: Youtube Lectures:
      • https://www.youtube.com/watch?v=JHrlF10v2Og&list=PL_iWQOsE6TfX7MaC6C3HcdOf1g337dlC9

  • Harjatin: track record. Open Source everything you make: github, blog post;