后验坍缩 | ViLab https://vilab.team/tag/%E5%90%8E%E9%AA%8C%E5%9D%8D%E7%BC%A9/ 后验坍缩 Hugo Blox Builder (https://hugoblox.com)en-usFri, 01 Jan 2021 00:00:00 +0000 https://vilab.team/media/icon_hu2896232876136423579.png 后验坍缩 https://vilab.team/tag/%E5%90%8E%E9%AA%8C%E5%9D%8D%E7%BC%A9/ VAE^ 2: Preventing Posterior Collapse of Variational Video Predictions in the Wild https://vilab.team/publication/vae-2-preventing-posterior-collapse-of-variational-video-pr/ Fri, 01 Jan 2021 00:00:00 +0000 https://vilab.team/publication/vae-2-preventing-posterior-collapse-of-variational-video-pr/ <p>本文针对野外视频预测中变分模型的后验坍缩问题，提出VAE^2方法。通过设计有效的潜在变量建模与训练机制，避免解码器忽略潜在码，从而增强模型对复杂动态场景的表征能力，提升视频预测的长期一致性与运动建模精度。实验表明该方法在多个视频预测基准上取得了优越性能。</p>