Legonet

Adapting Neural Architectures Between Domains is accepted by NeurIPS 2020. The power of deep neural networks is to be unleashed for analyzing a large volume of data (e.g. ImageNet), but the architecture search is often executed on another smaller dataset (e.g. CIFAR-10) to finish it in a feasible time. To enhance the generalization of the searched... neural architecture, why not taking a small piece of that large dataset into the process of NAS? Scientific Control for Reliable Neural Network Pruning is accepted by NeurIPS 2020. We prune redundant filters to achieve a compact deep neural network. But whether the filters you have pruned are really redundant? Let's set up a scientific control group for a more rigorous research design, so that the effect of all factors except the association between the filter and expected network output can be minimized. Learning Disentangled Representations with Latent Variation Predictability [code] is accepted by ECCV 2020. Latent traversal is popular to visualize the disentangled latent representations. Given variations in a single unit of the latent representation, it is expected that there is only a change in a single factor of variation of the data, but this impressive experimental observation is rarely explicitly encoded by existing methods. This paper defines the variation predictability to address this need. Optical Flow Distillation: Towards Efficient and Stable Video Style Transfer is accepted by ECCV 2020. In video style transfer, optical flow is widely adopted to boost the transfer stability, but its high computational complexity has been criticized for a long term. If we can formulate how optical flow helps in improving the transfer stability, it would be possible to graft it on another lightweight lightweight style transfer network that has no explicit input of the optical flow. Neural Architecture Search in a Proxy Validation Loss Landscape [code] is accepted by ICML 2020. Intermediate validation results in neural architecture search are invaluable but have rarely ...Read More