Convolutional Neural Networks (CNNs) currently represent the best tool for classification of image content. Over the past few years, this area of research brought great progress to image classification. One of the most significant breakthroughs in the beginning is the reliant classification of handwritten postal zip numbers and later the recognition of faces or license plates. Current state of the art applications use powerful real-time-capable networks that are able to detect multiple classes in images for detecting pedestrians, vehicles, obstacles and traffic signs in real-time. State of the art networks are ranked according to their overall ability to classify image content. At training time, the backpropagation algorithm is applied to the network in order to reduce the loss that the network is producing while classifying input images. This is done by increasing and decreasing filter weights that extract features from the given input gradually. During this process, one or more filter weights might develop same or similar values. In this case, only one of those filters is required and the other ones are redundant. By resetting the redundant filters, the network is possibly able to distinguish a different feature. This feature is then helping to distinguish between the classes and thus reducing the loss of the classification.
Mr. Zhang is presented with the task of examining the consequences of the adjustment on the weights during training and its effects on the classification results of a given CNN. The adjustment should be applied at different depths of the CNN, one layer at a time. While resetting filters, the output of an adjusted layer has to be adapted as well. Also, dropout should be deactivated since it might influence the overall effect of the weight adaption. The main part of the thesis consists of implementing required adaptions to the CNN, identifying the filter that should be reinitialized and tracking the influence of the weight adaption. Afterwards, different combinations and the impact of the depth of the layer that gets its weight adapted should be analyzed. Finally, well-founded decisions on the usage of weight adaption should be given.
The current state of the art shall be determined by conducting a literature research. The thesis shall include a well-documented presentation of the results; any source code created shall include sufficient annotation.