Model

Machine Learning

VC dimension
PAC

Performance measures

TP, FP, TN, FN
precision, recall, accuracy
AP, mAP
why threre are so many performance measures, such as AP, mAP, MR etc.

Representation

lr vs svm vs decision tree vs dnn
assumption
overfitting vs underfitting
bias vs variance
ensemble methods(bagging, boosting and stacking)
imbalance between classes
cross validation

Evaluation

loss
cross entropy loss and softmax(how to implement)
cross entropy loss vs mse

Optimization

first-order optimization or second-order optimization algorithm
optimization strategy, sgd, momentum, rmsprop, adam, adamw

Deep Learning

data augmentation

how to choose the data augmentation method

General

difference between all kinds of norm(in, ln, bn, gn, bn+gn)
how to get better results for bn besides using larger batch size
bn vs se
alpha in bn, how to prune network using alpha
bn vs whitening
eps and momentum in bn
loss normalization: use batch-wise norm vs sample-wise norm or others
interpreting confidence scores: process each class separately or not
the use of activation function
relu vs sigmoid
backpropagation
class imbalance
vanishing gradient problem

Convolutional neural networks

assumption of convolutional neural network
why convolutional layer not fc layer
features fusion methods
dilated conv vs deconvolution
receptive field calculation
why we don't need bias in conv when networks are pluged in bn
backpropagation of pooling layer and bn
exponential moving average
translation invariance and translation equalvariance
the implementation of dilated conv in tf
why no bn in fc layer
resnetv1 -> resnetv2 (preactivation)
how to improve downsample block of resnet
how to design an efficient network(perspective of hardware and architecture, head, tail, body, block, downsample module)
features of efficient network architecture
classifier to detector
reorg(space2conv) implemented by conv

Object detection

one stage vs two stage
ssd vs yolo vs retinanet
roi align vs roi pooling
anchor matching strategy
positive, negative, ignore anchor
objective function
how to detect small objects
how to get better detection performance
how to get faster detection model
why multi-scale and how
data augmentation
train from scratch
freeze part of layers or not
the difference between face detection and pedestrain detection
roi align -> roiconv
why anchor free
SNIP, tridentnet

Programming basics

leetcode
copy vs deepcopy in python
*((*float)(&(int a = 2;)))
0.35 / 0.05

Linux basics

command
tool chain, include, library, environment variable