Computer Vision News - April 2019

Project Management Tip 13 Computer Vision News In deep learning, once your data is well organized, you have done a big slice of the work! reality. When the project is ongoing, you can sometimes guide data collection about the data that is required. This is a huge advantage, since the major condition for the network’s success is a varied and representative dataset. If you are lucky enough to direct the data collection, you could point out the faults of the data. When you retrain your network with the improved data, you can now analyze the resulting errors, and better understand the network’s faults. You could ask for the data collection’s help in supplying more examples of those peculiar cases in which the network fails. Here, it’s important to go into details and not content ourselves with the numeric results of the network’s grade. A project manager could easily be excited about a high scoring network. He believes it to be very well trained - while in reality, validation set was limited and degenerated, and perhaps it did not evolve together with the ever growing and more complex training set. It could also happen that the validation dataset is varied, but the measurement you use to grade your network is not well representing its achievement, now that the dataset has changed. Remember to verify, by sampling and inspecting actual examples, that your validation dataset and its quantitative evaluation are still informative and reliable. Assessing improvements with an ever growing data require a system of code that will allow a very clear follow up of the launched trainings. Comparing network achievements makes sense only if both networks were trained on an equally complicated dataset. We’d like to avoid the unfortunate moment of staring a list of weights files, without having any clue about what do their quantitative scores mean and what is the conclusion of all this. For this sake, I recommend to keep for each weights file a clear follow up of the training process it went through, linking it or saving it in a directory with all the relevant information: its network architecture, network hyperparameters, a list of data items it was trained on, a list of data items it was evaluated on and its score. Finally, when your dataset is quite mature, and you inspect the errors and ask yourself: what is the story behind these errors? You may find out that the network has some difficulty in understanding the core question you are asking it, despite the varied data. At this point, you might consider changing the question you are asking, which might entail modification of the network’s interface and thus require some bravery. These are exciting moments in a project, since they invite you to see the problem with new eyes. They hold within them the opportunity for a step forward in the project’s achievements. More articles on Project Management Management When data keeps coming…