A conventional image recognition problem often by the divide and conquer its partial decomposition preprocessing, feature extraction and selection, classifier design several steps. Motivation partition method is to decompose a master image recognition problem simple and clear controllability several small sub-problems. But when the child step by step to solve the problem, despite the availability of sub-optimal solutions on the issue, but sub-optimal solution on the issue does not mean you get the final solution of the global problem.
Depth study provides an "end to end" learning paradigm, the whole learning process is not artificially divided into sub-problems, but entirely to deep learning learning model directly from the raw data to the desired output mapping.
As shown, the depth of the model for which the input data is in the form of any person, without the original sample processed, then the subsequent layer is stacked on the input operation of many layers. These operations entire layer can be regarded as a complex function Fcnn, the final loss function by the positive data loss (data loss) and the model parameter is of loss (regularization loss) composed, depth training model is the ultimate loss of the driving case of and updating the model parameters to a back propagation network layers. Training model may simply be abstracted from the raw data directly fitted to the end of the target, and an intermediate member which acts as the original data in the map is then mapped characterized labeled sample.
To sum up: end-learning is actually not any additional processing, from the raw data input to the output task results, and predict the whole training process, it is done in the model of.
Original: https://blog.csdn.net/alawaka2018/article/details/80388808
Depth understanding of end2end study of
neural networks, often see end2end training methods.
end2end There are different scenarios in different interpretations, for the visual field, enter the neural network as a picture, the output is the result we want. E.g:
1, autopilot, enter the path to the picture, the output is the steering angle.
2, robot vision, is a picture input and output is the command robot movement.
3, target recognition, an input image to be detected, the output is detected and the target position of the target category.
From the perspective of the detection target understood:
Carried out under non-end2end mode detection, general neural network will be divided into two parts, the first part is the Region Proposal, the function of this part is responsible for identifying the location and size of the picture goals. The second part is to determine whether it is the presence of the target, and the target is the first portion of which type, is typically determined in CNN accomplished. Such routines are generally often used in RCNN in.
In end2end mode, all addressed by the location and category of a model, the original image input, the output results of the model is the location and category. One end of the original input image, the output end of the desired result, only care about input and output, regardless of all the intermediate steps.
end2end benefits:
By reducing the artificial pretreatment and subsequent processing, as much as possible from the original input model to the final output, to more models can be automatically adjusted according to the data space, increase the overall fit of the model. end2end it emphasized that the global optimum, the middle part of the local optimum and can not represent the overall best.
Original: https: //blog.csdn.net/qq_17278169/article/details/77337308