Computer Vision News - April 2018

This month we publish the second part of Project Management in Deep Learning. The lecture is given by Yehiel Shilo , RSIP Vision. Yehiel earned a PhD from the Hebrew University of Jerusalem. The first part is here . It’s another tip by RSIP Vision about Project Management in Computer Vision . We learned last month how to prepare and manage data . In the second part of this lecture we deal with the model itself. The first step is to properly define the problem at hand and the objective of the work. These, together with the criteria by which we shall measure success, should be clear and precise. That will help to know at any moment whether the project is succeeding or not. Then comes the time to think at the architecture of the project. The project manager shall chose the best among the many architectures offered in the literature, as their number increases almost on a daily base. This might be a very specific architecture or a generalist one which is fit to solve complex problems. This process may take time, as it is key to consult the latest findings made available. The chosen model should be tested using the existing data to check that it works properly. The next step is training the model over a number of epochs. The purpose is to understand how the model deals with and reacts to the training data. Is it able to define the appropriate weights for the network, so that it solves the problem at least over the training data. If it works, then fine. If it doesn’t, one option is to expand the dataset used. It is also advisable to consider again whether the chosen architecture is fit for the problem at hand. When the model is successful in understanding the training data, it is time to bring in the previously prepared validation data. When this is successful too, that’s very good news. When it’s not, the chief option is always to add more data. But we can also ask ourselves: did we train the model enough? Shouldn’t we add more epochs? At this point, it is also recommended to compare training data and validation data, to discover why only the former are explained by the chosen architecture. A brief setback can be a great opportunity to think again. Once some of the hyper-parameters have been changed and validation works, we apply the model to test data. At this point, hyper-parameters are not amendable. When test data gives results in line with the success measures defined at the beginning, the project is achieved. Otherwise, expanding data is always a solution, especially when results are very close to requirements. It is also worth checking again whether data was divided correctly between learning, validation and test, since only the latter does not work under the chosen architecture. Project Management in Deep Learning - part 2 Management Project Management Tip 25 Computer Vision News