【中文字幕】2017春季CS231n 斯坦福深度视觉识别课

17:19  z在x方向上的梯度是y

cifar10

concept : receptive field

• pooling layer:

Max pooling:

max pool with 2*2 filters and stride 2

(no overlap)

Generally, in pooling layer we don't use zero padding.

common setting:

filter size : 2/3 ; stride : 2

• Fully-connected layer

convolution layer

The factors that we should consider:

1. the size of filters
2. stride
3. how many filters you have

multiple filters - for example the number of filters is n, and then we will gain n activation maps.

(That means each of the filter is producing an activation maps.)

The filters at the earlier layers ->Low-level features:  like edges.

mid-level features: more complex features , like corners and blobs(斑点) and so on,

high-level features： more ensemble concepts than blobs.

[From simple to complex features.]

Assume we have 7*7 input and 3*3 filter , and then we will gain an 5*5 activation map.

Assume we have 7*7 input and 3*3 filter , and the stride is 2 . Then we will gain an 5*5 activation map.

Then we can think about it, how to compute the output size?

input ——N.  The size of filter——F.

e.g. N=7，F=3

stride 1=> (7-3)/1+1 =5

stride 2=> (7-3)/2+1 = 3

stride 3=> (7-3)/3+1 = 2.33 (error)

In practice: common to zero pad the border

pad——M

(N+M*2-F)/stride +1

parameters

note: each filter has one bias

input: 32*32*3

filters: 10 5*5

(5*5*3+1)*10 = 760

the application of ConvNets:

1. image retrieve
2. detection and location (self-driving cars...)
3. face-recognition
4. pose recognition
5. game playing: Alpha Go
6. image captioning
7. artwork and calligraphy

fully-connected neural networks(conditional neural networks)

convolutional neural networks（部分连接与权值共享）

distance metics   and values of K

The critical steps of the development of object recognition:

1.image segmentation

Reducing the complex structure of the object to simple structure

2.face detection

ML：SVM、boosting

Adaboost——face detection

3.feature matching

SIFT feature-The idea is that to match 

some features tend to remain diagnostic and invariant to changes(某些特征能够在变化中具有表现性和不变性)

the task of object recognition is identifying these critical features(关键的特征) on the object and then match the features to a similar object

4.recognize holistic scenes——spatial pyramid matching(algorithm)

There are features in the images that can give us clues about which type of scene it is.

5.object recognition

The comman data set of deep learning:

1.MNIST

2.PASCAL Visual Object Challenge

3.ImageNet

4.COCO

Neural Networks

Backpropagation

computational graph —— Chain rule

add gate

max gate

mul gate：交换