Perception wild concept is particularly important, is working to understand and diagnose network CNN, in which the image of a wild-aware than neurons do not have an impact on the value of neurons, so to ensure that all relevant neurons covered this image the region is very important;
the task requires a single pixel in the output image is predicted, so that each output pixel has a relatively large field of perception is very important when doing pre-test, each of the key information will not be omission.
A method of increasing the perceived field: theoretically be achieved by building a network of more layers increases the perceived linear domain, the convolution of the filter against the increase; method may be used in sampling, pooling, increase the perceived region, the current usually a combination of the two technologies;
The authors found that not all of the pixels in a perceptual domain diagram for output units have the same contribution: intuitive, the perceived influence of the middle of the field will have more pixels for output.
Forward propagation, the perceptual field of the intermediate pixel information can be transmitted to the output by a number of different paths, the edge pixel is relatively small. This has resulted in the reverse direction, coming through these paths gradient, such that the intermediate pixel gradient magnitude greater update
Wild perceived impact of the distribution is Gaussian distribution, the authors found an effective part of the field theory of perception is actually very small, because the Gaussian distribution from the middle decays very fast
Determine how many pixels the input field Perception affects the output neurons;
there is a conclusion: do not use the pool and down-sampling method in the residual network, along with training and effective in improving the perception wild range, while real-time sensing has become larger than the size of the wild entire size of the entire image field is valid or not sensing area covers the entire image.
In the residual network architecture model, the subsampling technique, theory of perception domain increase is very large, but also very effective domain-aware small;
The impact of the Gaussian distribution effectively reduce the perceived effective domain of
the heavy weights of 1 to manipulate the right convolution kernel initialization value smaller centers, external weights greater
optimization to maximize effective w perception of the size of the field
to solve the optimization problem , you get such a solution: the weight evenly distributed on the four corners of the convolution kernel, while the rest are zero.
Some initialization method to obtain such a distribution can improve the overall speed
From a structural point of view think CNN is a good effective measures to increase the perception domain, such dilate conv, skip-connection makes the perception field even smaller, dropout does not change the effective perception wild