Convolutional Neural Network of Supervision group 1

Convolutional Neural Network of «Supervision group» 1

2

Feedforward convolutional network of “Supervision group” 3

Feedforward neural networks Input 4

Convolutional filtering Feedforward Feeedforward Input 5

Convolutional network of Kunihiko Fukushima Convolutional filtering Image structural synthesis Examples of image matching 6

Schematic structure of eye In such way the physiologists test the outputs of neural cells crystalline lens cornea vitreous body retina probing rod of oscillograph Neural layers of retina pigment cells rod cells conus cells amacrinal cells Neuron cells and interneuronic dendrites interactions horizontal cells bipolar cells ganglionic cells core membrane axon synapse 7

A Complex cells Simple cells 8

9

Directed lighter Small mirror Projected mark of attention field 10

Weight 0 0 Y X X Information function of Muchnik and Zavalishin Positions of the extremes of local difference 11

Weight Weigh t 0 X 0 0 Y Y 0 X Synaptic weights having the shape of “Mexican hat” Weigh t 0 X X X Y 0 X Information functions proposed by Muchnik and Zavalishin 12

Вес X 0 Paraboloid of revolution Matching and masking the subimage inside the AF Scanning the image with an attention field of decreasing size Separating an object from background, measuring the object position and elongation parameters Adapting the position and shape of attention field 13

Neuro-physiological texture description after David Marr 1. Mean local brightness. 2. Local spatial density of distribution of texels. 3. Average sizes (length and width) of texels. 4. Spatial orientation of texels. 5. Distance between neighboring similar texels connected with permissible straight line. 6. Orientation of permissible straight line connecting similar texels. Bela Yulesz surmised that there are special texel detectors in the living vision systems of the shape of “Mexican hat” proposed by David Marr. 14 14

Texture structural description by convolutional neural network of Lutsiv and C° Weight 0 х Texels presented as oriented elongated micro-objects Aerial photograph Detectors of texels after Lutsiv Weight Detectors of texels after Vistness Historgam of texel types Weight Feature vector х Detectors of texel borders after Vistnes Clustering of texture description feature vectors Result of image segmentation by texture features 15 15

Training of Supervision’s deep convolutional neural network Examples of training images 16

Training the Supervision’s deep convolutional neural network Too deep networks encounter difficulties in training by backward propagation of error signal Backward pass Backward propagation of error signal 17

Pretraining of weights by restricted Boltzmann machine 18

Energy of state of Boltzmann machine Training of Boltzmann machine Probability density of state of i-th neuron 19

Energy of state of Boltzmann machine Probability density of state of i-th neuron 20

Training of Boltzmann machine 21

Energy of state of restricted Boltzmann machine Training of restricted Boltzmann machine , where 22

Energy of state of restricted Boltzmann machine , where 23

Training of restricted Boltzmann machine 24

Training acceleration by changing of sigmoid (logistic, activation) function 25

26

The 224 256 224 27

Training of Supervision’s deep convolutional neural network 48+48 28

29

Image classification in Supervision’s deep convolutional neural network max 30

Images classified as belonging to same classes 31

32

Is 33

Any questions? 34