Computer Vision News - January 2017

Our CEO Ron Soferman has launched a series of lectures to provide a robust yet simple overview of how to ensure that computer vision projects respect goals, budget and deadlines. This month we learn about Managing Deep Learning projects . Nowadays, an overwhelming portion of computer vision scientific articles and industrial projects are based on Deep Learning technology . This fact by itself would be sufficient to justify that all deep learning projects be planned and conducted in the most educated way. Another reason is that managing a “black box” project is quite challenging: deep learning models function even if no one can specifically explain how they work and why some of them achieve the desired results and some don’t. This is hardly acceptable for a project manager who wants to predict where the work is heading, what are its chances of success, what are the risks and how to minimize them. We will now clarify two modes of work in this area and two management approaches. Some managers are very knowledgeable in deep learning, while others have only a limited experience; the same distinction could be done with the engineers working in their teams. In most of the cases, you want initial results to confirm the validity of the model following a first reasonable effort on simple networks. Most of the cases will involve transfer learning using popular networks like AlexNet , GoogLeNet , VGG-16 and the like; another level of training like SVM (Support Vector Machines) can adapt the network to the specific task. Even when initial results are below expectations, they will nonetheless show encouraging signs of the validity of the model, enabling to reduce the risks. Such a project is not expected to take more than one month, a rough figure depending on the data and on the specific traits of the task at hand. When the team members are not well versed on deep learning, there is no use in huge investments in data, training and data acquisition, ground truth on thousands of images and training networks for days and weeks. Thus, it is the project manager's role to get initial results at the first stages. There will be many ways to improve them later. Results may fall short of expectations for several reasons and the very popular choice of running to gather additional data is not the optimal solution, since a reasonably-sized dataset should give the necessary indications, especially when engineers working on the project are skilled and informed: sometimes good adjustments bring better results than expanded datasets. For example: data augmentation designed for the task at hand and network parameters adjustment, such as learning rate and momentum, can improve the performance significantly. From our experience at RSIP Vision , these two aspects - risk minimization and gaining confidence in the validity of the model - are crucial in the attempt to solve the problem in the best way. “Two crucial steps: risk minimization and gaining confidence in the validity of the model” 29 Computer Vision News Project Management Tip How to Manage Projects in Deep Learning Management