See How a Neural Net Works

Visualize the inner working of a trained image classification network

Load a pretrained network:

In[1]:=

net=NetModel["Inception V1 Trained on ImageNet Competition Data"]

Out[1]=

NetChain



Input port:	image
Output port:	class



The image features of the neurons in the first convolution layer are simply given by their convolution kernels:

In[2]:=

kernels=NetExtract[net,{"conv1_7x7_s2","Weights"}];

In[3]:=

Map[ImageAdjust@Image[#,InterleavingFalse]&,Normal@kernels]

Out[3]=





Specify a layer and feature to visualize:

In[4]:=

layer="inception_4b";channel=56;

To obtain an impression of what a neuron encodes, truncate the network at the given layer and attach layers that extract the total square activity of the specified neurons:

In[5]:=

featureNet=NetAppend[NetTake[net,layer],"L2_norm"{PartLayer[{channel}],ElementwiseLayer[x

],SummationLayer[]}]

Out[5]=

NetChain



Input port:	image
Output port:	real



Create an initial image that exhibits randomness at all scales:

In[6]:=

meanColor=NetExtract[net,{"Input","MeanImage"}];netDims=NetExtract[net,{"Input","ImageSize"}];

In[7]:=

img=Nest[imgImageEffect[GaussianFilter[img,{6,2},Padding"Periodic"],{"GaussianNoise",0.05}],ConstantImage[meanColor,netDims],8]

Out[7]=

The

NetPortGradient

["Input"] command provides the backpropagation gradient at the input port of our network. The following function normalizes the maximum gradient coefficient to 1/8 and converts it into an image:

In[8]:=

gradientNormalize=Image[Normalize[Normal@featureNet[#,NetPortGradient["Input"]],array8Max@Abs@Flatten[array]],InterleavingFalse]&;

Before iteratively updating our image with the above gradient image, we need to introduce a jitter function which shifts an image by a small offset:

In[9]:=

jitter[{dx_,dy_}][img_Image]:=ImagePad[img,{dx*{1,-1},dy*{1,-1}},"Periodic"]

Incorporating a random jitter into the following loop helps to spread out local artifacts of the backpropagation and promotes a more regularized image:

In[10]:=

list=NestList[With[{offset=RandomInteger[{-8,8},{2}]},ImageClip@ImageAdd[#,jitter[-offset][gradientNormalize[jitter[offset][#]]]]]&,img,265];

Updating the initial, random image 256 times returns the following image. The neurons in question seem to encode small birds or chickens:

In[11]:=

ImageResize[Last[list],150]

Out[11]=

See the evolution of the neurons:

In[12]:=

FrameListVideo[list]

Out[12]=

Publisher Information

Contributed by: Wolfram Staff

Wolfram Language Example Repository

See How a Neural Net Works

See Also

Publisher Information

See How a Neural Net Works

See Also

Related Symbols

Publisher Information