2018-12 在职刷题保持状态， 每天1题+ML insights

Day69

Solved 525

Solved 572

Solved 314

Day70

Solved 523

Solved 554
Solved 489

Finally root caused a very low-level bug related to MMX instruction and FPU registers.

ML Reviewed:

DNN

Weight initialization (Xavier method to render the variance to 1/n, n is the input dimension, help gradient explosion/diminishing)
regularization (l1,l2,dropout, early stopping)

Planned review of optimization today.

ML Reviewed:

Momentum, RMSProp, Adam. Bias corrections.
Batch norm (affine function of normalized Z, which helps alleviate covariate shift. Note that need to have the estimates of mu and sigma in order to apply batch norm during test time)

Day71

Solved 523

Solved 554

Solved 489

ML Reviewed:

Momentum, RMSProp, Adam. Bias corrections.

Day72

Solved 477

Solved 461

Solved 404
ML Reviewed:

Numpy array,arithmetic operations.

Convolutional NN.
Input preparation:
Padding, Strided CNN, Filter/Kernel and dims. Convoluation over volumes.

Day73

Solved 398

Solved 377

Solved 334
ML Reviewed:
1X1 Conv. Pooling.

Classic CNNs:
LeNet,AlexNet,VGG-16,ResNet(residual block), Inception Net.

Day74

Solved 286

Solved 275

Solved 957

Day75

Solved 899

Solved 251

Solved 890

Day76

Solved 443

Solved 851

Solved 842

Day77

Solved 362

Solved 755 with monotonic stack
Solved 735 with stack.

ML reviewed:

Structured review of the following topics, alg, pros/cons, pratical use tips:
Decision Tree (ID3, CART, Gini purity)
Random forest over bootstrapped samples
Adaboosting
Gradient boosting decision Tree
K-means (Forgy and random initialization, K-means ++ initialization)
Naive Bayes (Gaussian, Multinomial, Bernoulli)
SVM

Tech Blog reading:

Linkedin Pro-ML platform

Generalized Linear Mixture Model

Linkedin Feed

Day78

Solved: binary random generator [0,1] to generate uniform distribution in range [0, 6]

Solved: biased random generator [0,1] to generate uniform distribution in range [0, 6]
Solved 445.


ML reviwed:

Recurrent Neural Network (gradient exploding/vanishing)

LSTM
Interviewed with Linkedin today. Finger crossed.