This is just a post for myself to write notes while watching videos, so it may contain lot of typos and some mistakes.
- Understanding basic programming
- probability basics: random variable,
- basic linear algebra: matrix, product, eigen vector
Aim: To do an awesome project by the end of project and gain basics useful forever.
I am already familar with basics of programming, in case of probability, I know about bayes theoreum, yet there are more topics which I don’t, so cheatsheets for now. I am not familar with linear algebra much now, yet I remeber taking Rachel lectures to learn about SVD, NMF etc.
In the less Andrew covered following topics:
- Difference b/w CS229, CS230, CS229A?
- History(Arthur Samuel, chess program)
- Deep learning(automated car moving)
eg: given age, tumour size - 2 feature input, predict if tumour is malignant(classif)
Andrew started off by theoratically representing, linear regression. He explained about gradient descent, and talked about the ambiguity of big enough data now(~1-10 million datapoints now). In case of a batch gradient descent, we use entire training data fully, even for 1 parameter update.
Some mathematical properites:
∇A means derivate of a
trace AB = trace BA
tr ABC = tr CAB
cost = 0.5*np.sum(A[i] - y[i])**2
Proofs are clearly explained in the lecture notes. In end, he showed derivation of Newton equation which reaches the optimal convergence point in a single update(only for linear reg). The contour and loss curver visualisation to show effect of learning was awesome.
Andrew talks about locally weighted regression method where, each points is fitted in a straight depending on it’s curve and dependance. It used for low-dim data with less features and is an example of non-parametric learing algorithm.
- why least square formula?
- maximum likelehod
Read: Why understanding backprop is necessary?
Lesson7 (Kernels in SVM)
- use optimisation problem
Representation theoreum proof, ie
|wb||^2-> denoting kernels with objective function|
can’t understand: dual optimization, convex optimization, etc
ie Kernel trick: Kernel function(x,z) = Phi(x)Transpose* Phi(z)
ie K(z,z) = (x T z)^2