[ Archived Post ] Lecture 1 | Syllabus + Introduction + Basics of Linear Algebra (Hopkins)

Please note that this post is for my own educational purpose.

Unsupervised learning → discovers structure in the high dimensional dataset. (starting with linear subspace → and more → we can create things more robust). (we are going to start from PCA).

Some of the models → are generative models. (GPCA → generalized PCA). (there are going to be two midterms and more) → very interesting perspective.

More like the introduction to the course → and how the gradings are happening.

A good percent of the lecture will be recorded → THANK YOU!

A high dimensional data → video image and audio → biomedical data → a large amount of data exist. (very cool).

A lot of images are taken and more → electronic and cameras. (started from 2010). (facebook → another giant source of data). (Youtube and more → need to process these data abuses the structure). (labels and learned task → this is important).

Unsupervised learning → learns the structure of the data. (Big Data → biomedical data → more and more data are being collected → we need a method to analyze this). (Data can define everything → and answer a lot of stuff).

At first, we can learn a linear subspace → , not a manifold → for face images → we can capture light images and more. (a lot of faces are being captured together). (or union of subspace is also possible).

The first method → PCA → very long history → in reality → data is very noisy → how can we fight these bias? (there is a subspace → but there is a lot of noise).

How about outliers → how can we deal with that → or points data have not been collected? (find out what missing entries are and more → this is a low rank approximation). (Robust PCA and a probabilistic PCA).

Non-linear sub-space → much harder problem

The general structure has to be captured → this is a hard problem. (this can actually happen in picture space) → non-linear transform. (projection of data into non-linear transform and more).

We are going to cover everything → super sexy. (GPCA → most powerful PCA method → find the subspace as well). (noise, outlier and missing data → all of them are problems). (non-convex optimization).

Tons and tons of application → very sexy. (motion segmentation and more). (image segmentation and more)

Generalized → PCA → some of them are iterative and some of them are generative models.

Linear algebra.

Some matrix is important → compared to other ones.

Positive Semi Definite matrix → another important → Trace.

Tr(A) → diag values → some of eigen values.

Other properties of the trace and → interesting results.

Very sexy properties.

https://jaedukseo.me I love to make my own notes my guy, let's get LIT with KNOWLEDGE in my GARAGE