Code available at https://github.com/mfederici/Multi-View-Information-Bottleneck 摘要:信息瓶颈原理为表示学习提供了一种信息论方法,通过训练编码器保留与预测标签相关的所有信息,同时最小化表示中其他多余信息的数量。然而,最初的公式需要标记数据来识别多余的信息。在这项工作中,我们将这种能力扩展到多视图无监督设置,其中提供
原摘要: Multi-view clustering aims to employ semantic information from multiple perspectives to accomplish the clustering task. However, a crucial concern in this domain is the selection of distinctive f
encoder p θ ( z 1 ∣ v 1 ) _θ(z_1|v_1) θ(z1∣v1),D S K L _{SKL} SKL represents the symmetrized KL divergence. I ˆ ξ ( z 1 ; z 2 ) \^I_ξ(z_1; z_2) Iˆξ(z1;z2) refers to the sample-based parametric