Increasing the performance of a CNN
How to improve the accuracy of a convolutional neural network?
There are several ways to do this, the first being to change the network parameters. Increasing the number of epochs within certain limits definitely leads to an increase in inaccuracy. This can also include changing the learning rate from higher to lower in the process of training the network or using optimization algorithms that implement this principle.
The second is data augmentation. This method is implemented through the use of a whole galaxy of techniques. Only a small percentage of them are presented on the slide, ranging from classical geometric transformations, using filters and image mixing, to using GAN networks that generate new images.
The third approach is based on changing the topology of the network, related to its depth and width. So, it is preferable to change the depth of the network, which leads to a linear increase in computational complexity, rather than its width.
Finally, the fourth approach is related to the effects of underfitting and overfitting.
Original article written by Rinat S.
