ICCV Daily 2019 - Tuesday

search or a bigger perspective of the problem called AutoML will be the future of artificial intelligence, computer vision and machine learning. However, he says there are still some challenges to be solved before AutoML can be applied to a wide range of applications. As well as speed and stability, the scale of architecture search – meaning that the architecture can be searched in an even bigger space – is an issue. Currently,many problems limit the architecture search in a relatively small space. This is not ideal but has to be done because when this constraint is freed, most architecture search methods become either too slow or less stable. Also, many people apply hardware constraints because the search network is not so friendly to hardware. Another challenge is the difference between the search stage and evaluation stage which means that they need to be two separate stages. In the future, Lingxi hopes that they can design a method that can search directly, and after finishing the search stage does not need to be retrained so that the model can be applied directly to real-world problems. Exploring solutions to these challenges will be future steps for this work. How can this work be practically applied in the real world? Xin tells us: “One of the applications is on the image classification task. Just to search a network for the classification model and when you get a picture, the model can immediately tell us what the object in the picture is.” In conclusion, Lingxi says neural architecture search and AutoML are very promising research subfields and will have a significant impact on artificial intelligence, computer vision and other fields such as computational linguistics. It is still in a preliminary stage and much work remains to be done, but he hopes that their work will inspire future researchers in the field. To learn more about Xin and Lingxi’s fascinating work, come along to their oral and poster today at 13:35 in Hall D1. 9 Xin Chen and Lingxi DA I L Y