# XSleepNet (2021) >[Source](https://arxiv.org/pdf/2007.05492.pdf) - ### Abstract - XSleepNet uses multiview inputs and takes in both raw signals and time frequency images. - It learns jointly on these inputs complementarily. - ### Model - The raw signal network is large and based on CNN, the time frequency signal network is compact and uses RNN. - They are trained such that learning on stream that is generalising is accepted and overfitting one is discouraged. - ![](https://i.imgur.com/SKMMy9y.png) - Architecture used bidirectional LSTMs. - It has 3 softmaz outputs, 2 are view specific and 3rd is joint output. - Loss weights are weighted on ration of generalization and overfitting. - w = (1/z) * (G/O<sup>2</sup>), Z is normalization factor. - Based on two approximations used for G and O, XsleepNet 1 and XsleepNet 2 are architectures used which employ loss and tangent curves respectively. - ### Evaluation - Done of 5 datasets sleep-edf20,sleep-edf78,MASS,SHHS,PhysioNet2018. - Has an accuracy about 80-85 in all datasets with XsleepNet2 performing better.