- Level-DAE: Denoising Autoencoders for Self-supervised Level Cloud Studying(arXiv)
Creator : Yabin Zhang, Jiehong Lin, Ruihuang Li, Kui Jia, Lei Zhang
Summary : Masked autoencoder has demonstrated its effectiveness in self-supervised level cloud studying. Contemplating that masking is a type of corruption, on this work we discover a extra common denoising autoencoder for level cloud studying (Level-DAE) by investigating extra forms of corruptions past masking. Particularly, we degrade the purpose cloud with sure corruptions as enter, and be taught an encoder-decoder mannequin to reconstruct the unique level cloud from its corrupted model. Three corruption households (ie, density/masking, noise, and affine transformation) and a complete of fourteen corruption sorts are investigated with conventional non-Transformer encoders. In addition to the favored masking corruption, we determine one other efficient corruption household, ie, affine transformation. The affine transformation disturbs all factors globally, which is complementary to the masking corruption the place some native areas are dropped. We additionally validate the effectiveness of affine transformation corruption with the Transformer backbones, the place we decompose the reconstruction of the whole level cloud into the reconstructions of detailed native patches and tough international form, assuaging the place leakage downside within the reconstruction. In depth experiments on duties of object classification, few-shot studying, robustness testing, half segmentation, and 3D object detection validate the effectiveness of the proposed technique. The codes can be found at url{https://github.com/YBZh/Level-DAE}
2. Illustration Studying primarily based and Interpretable Reactor System Analysis Utilizing Denoising Padded Autoencoder(arXiv)
Creator : Chengyuan Li, Zhifang Qiu, Zhangrui Yan, Meifu Li
Summary : With the mass development of Gen III nuclear reactors, it’s a common development to make use of deep studying (DL) methods for quick and efficient prognosis of attainable accidents. To beat the frequent issues of earlier work in diagnosing reactor accidents utilizing deep studying principle, this paper proposes a diagnostic course of that ensures robustness to noisy and crippled information and is interpretable. First, a novel Denoising Padded Autoencoder (DPAE) is proposed for illustration extraction of monitoring information, with illustration extractor nonetheless efficient on disturbed information with signal-to-noise ratios as much as 25.0 and monitoring information lacking as much as 40.0%. Secondly, a diagnostic framework utilizing DPAE encoder for extraction of representations adopted by shallow statistical studying algorithms is proposed, and such stepwise diagnostic strategy is examined on disturbed datasets with 41.8% and 80.8% larger classification and regression job analysis metrics, as compared with the end-to-end diagnostic approaches. Lastly, a hierarchical interpretation algorithm utilizing SHAP and have ablation is introduced to research the significance of the enter monitoring parameters and validate the effectiveness of the excessive significance parameters. The outcomes of this examine present a referential technique for constructing sturdy and interpretable clever reactor anomaly prognosis techniques in eventualities with excessive security necessities. △ Much less