Just saying my random thoughts, as if someone's reading!
Yesterday's break was very effective because I felt energized throughout today! - so much that I decided to extend my stream to 15 hours. Of course, I took a lot of breaks and did other tasks other than studying throughout the day, but overall today was my most successful stream ever. Also gained a new subscriber and studied with them for several hours - was super fun. My studies consisted of ~9 hours of Machine Learning/Optimization and ~3 hours of Analysis. I'm stuck at a pretty gnarly section in my ML course (Ridge Regression & LASSO) which requires me to extensively cross reference an Optimization class I took in the past (evidently, I didn't learn the content well enough). I always try to learn a concept to the point that I'd feel comfortable lecturing the topic to a small group. Unfortunately, many of the concepts here are quite high level and taught using intuition, so it's taking me a long time to learn it.
Didn't stream today. Was pretty burned out after my three consecutive full streams. I'm getting better at not ending the stream early, but I still need to work on being present and ready for each session. I spent the day calling friends, family, and ended it with a nice dinner at the school cafeteria.
Today I took a break from streaming.
Today I reviewed some linear algebra results to do some problems from homework. I was stuck in a Wikipedia rabbit hole on matrix norms and positive semidefiniteness for a few hours. I returned home and ended the stream early. Had some great Indian food in Fremont with my brother, thanks to Arnav's recommendation.
I also learned that I was hired as course staff for Berkeley's CS 70, which is a class on Discrete Math and Probability. I'm very excited to help facilitate it over the summer CS 70 is a notorious class among students here due to its pace, so I want to do a great job in helping students understand the concepts. I had also spent the better part of last year tutoring this class individually and in small groups, so I'm grateful that my experience was recognized by the University today.
Today was all about Multivariate Gaussians. I read two notes on the topic with two different interpretations. The first paper focused on the geometric construction of Multivariate Gaussians with a defined density function. We can then think of the density function as "de-sphering" a standard Multivariate Gaussian into its transformed self. This perspective becomes even more clear when you think about the elliptical isocontours of the density. We orthonormally diagonalize the covariance matrix into simple scaling, rotation, and flipping operations and apply them to a sphere until it becomes the hyperellipsoid in the actual distribution.
The other note I read proved many theorems that the first note used liberally, such as the distribution of an affine transformation of the standard Multivariate Gaussian. It also touched on other interesting propertiess, such as how the MMSE estimator is actually the LLSE estimator. It was satisfying to see the two notes complement each other so well.
Later in the day, I tried to tackle some Topology. I reviewed some basic definitions in Metric Spaces and did some exercises in the book. I was feeling very tired at this point and had to take a few 10 minute naps mid session. I also haven't been eating well, though that's entirely due to my stubbornness to cook despite my lack of aptitude for it. I think I will take Saturday off in order to recharge and figure out how to cook a really nice dish.
Today is also the second time that I've hit the full 12 hours of streaming, and I'm very satisfied by that.
Wrapped up my exploration of Information Theory for now. I learned about Joint Entropy, Conditional Entropy, Mutual Information, and KL Divergence. KL Divergence was especially interesting. Although it's not a valid metric, the similar Jensen-Shannon divergence is. There's more content here that I want to learn about later, like Binary Erasure Channels and Huffman Codings.
I returned to studying ML, now having some familiarity in Info Theory. I learned a very cool proof on the equivalence of various loss functions. Minimizing the KL Divergence is the same as Maximizing the Cross Entropy, which is the same as Maximum Likelihood Estimate (this last equivalence is just the best, really). I'll write an article detailing the proof soon!
After lunch, I spent some time on Topology and promoting my stream on Discord. I met a very cool person there who shares a lot of mutual contacts.
Today was really hard. My laptop kept on overheating and my stream kept on crashing from some cryptic encoder error. I spent upwards of 6 hours in OBS nudging sliders and pondering over buttons to find the optimal parameters. Like a monkey doing iterative optimization without a gradient vector. Needless to say, I didn't do much studying.
Things brightened up towards the end though. I found a great setting for OBS - the stream looks and sounds better than ever and my laptop hardly breaks a sweat! I also graduated from my trusty phone timer by writing a Lua script for a Pomodoro Timer. It even beeps when time's up! You can see me working on the script in real time in my "Day 5 (5)" stream. Here's the repo.
The theme of the day was improving my stream, slowly but surely. Today got me thinking about how I want to grow my channel in the future. I don't have the beautiful skyline views or natural sights that other Studytubers have (my quaint shared apartment at Berkeley has no such features), but I do know that I'm studying a subject that a lot of people are interested in. I will try to make this into my niche in the saturated Study With Me space. I will also try to be more interactive/informative in the chat for all my viewers. I also should record videos. Currently I'm thinking about a 20 minute video about all the vector/matrix calculus needed for Machine Learning. Let me know if you have any other suggestions!
Finally managed to stream almost the entire 12 hours. Some things I learned: (1) My brain shuts off after lunch. Maybe I should try taking longer breaks, 10 minute naps, or a set of pull ups. The implication is (2) If I'm learning a concept for the first time, I should prioritize it for the morning instead of less intensive things like easy homework questions and review.
Today I began to tackle the extra content of EECS 126 that wasn't covered in my semester's iteration of the course. I learned about a more rigorous definition of Random Variables and some more Concentration Inequalities. Later in the day, I decided to dip my toes into Information Theory. Entropy is such a cool idea. Hope to learn more about it tomorrow.
The stream was a success overall, but I feel like I need to relax harder during my breaks so I can focus. It usually takes me a long time to read math, but even more so today. My one and only regular viewer showed up today in the comments once again and we talked about ML problems in Compbio. Cool stuff!
Got to compactness in Rudin Topology! I owe some thanks to Professor Anantharam a few semesters ago for cruelly putting topology questions on our linear algebra problem sets - so far the concepts have been pretty intuitive.
Today my stream got its first livestream comment ever! Had a short and wholesome conversation with a medical school student interested in learning more about AI. Shoutouts if you're reading this, and thanks for hopping on the stream :)))
Hello hello! Today I got through 2 notes from Berkeley's ML course and 2 lectures of Analysis. I'm excited to read some more Rudin Ch 2 tomorrow and learn some topology!
The stream was pretty good as well! We hit 3 concurrent viewers at one point (I swear it wasn't just me on different windows). I didn't quite reach the 12 hours mark, but I fixed some things with my streaming setup so we shud be chilling.