This morning, the ICLR 2018 paper acceptance results were announced, and we will take you to a general overview of this year's ICLR 2018 paper acceptance.
ICLR stands for International Conference of Learning Representation, which was jointly initiated by Lecun, Hinton and Bengio, three veterans of neural networks. In recent years, with the success of deep learning in engineering practice, the ICLR conference has also developed into the top conference of neural networks in just a few years.
Paper acceptance rate:
2.3% of oral presentations, 31.4% of posters accepted, 9% of workshops, and 51% of rejections.
ICLR Oral Presentation Paper Quick Facts:
More than half of the papers in the ICLR oral papers will become ICLR Best papers, which also represent the research direction in 2018. Let's briefly introduce this year's oral papers. Because the ICLR conference papers have a wide range and new directions, we Nor can it be comprehensive.
Wasserstein Auto-Encoders （Max Planck Institute）
This paper proposes to use the Wasserstein distance as a metric in the Variation Auto-Encoder, so that for the first time VAE can produce comparable results to the Generative Adversarial Network. And WAE theoretically links VAE and GAN. It is a rare good paper that combines both theory and practice. The image generated by WAE is as follows:
Spherical CNNs (Max Welling Group, University of Amsterdam)
Convolutional neural networks can only be used in 2D planar images, but in recent years, many problems such as robot motion and autonomous driving require analysis of spherical images. The traditional method is to project spherical image to 2D planar image, but this process will produce distortion, as shown below:
So the author proposed spherical CNN. Spherical CNN uses Fourier transform to avoid excessive computation. The schematic diagram of realizing spherical CNN through Fourier transform is as follows:
It is believed that the spherical CNN proposed in this paper can be widely used in the tasks of autonomous driving and robot motion.