coupling layerを特徴として持つFlow-based深層生成モデルです 8 NICE [8]ではAddictive coupling layerを採用し、各レイヤーの変換前後で体積が保存されるが、 (volume preserving)が、Affine coupling layerは非体積保存(non-volume preserving)で ヤコビアンの行列式はその対角成分の和として計算できる : transformation from i to j, : its Jacobian determinant concat Affine coupling layer x z S T x1 x2 z1 z2 ◉ + f f-1 p p-1 f f-1 f f-1 ・・・ ・・・ Permutation layer pを交互に挟み、全次元に対して変換可能なアーキテクチャを構成する Forward transformation Fxz Inverse transformation Fzx
Generative Flow with Invertible 1×1 Convolutions. arXiv preprint, arXiv:1807.03039v2 (2018). [2] J. Regier, et al., Celeste: Variational inference for a generative model of astronomical images. arXiv preprint, arXiv:1506.01351 (2015). [3] A. van den Oord, et al. WaveNet: A Generative Model for Raw Audio. arXiv preprint, arXiv:1609.03499 (2016). [4] R. Gómez-Bombarelli, et al. Automatic chemical design using a data-driven continuous representation of molecules. arXiv preprint, arXiv:1610.02415x3 (2016). [5] https://www.youtube.com/watch?v=JrO5fSskISY [6] http://www.shakirm.com/slides/DeepGenModelsTutorial.pdf [7] L. Dinh, L. Sohl-Dickstein, and S. Bengio, Density estimation using Real NVP. arXiv preprint, arXiv:1605.08803 (2016). [8] L. Dinh, D. Krueger, and Y. Bengio, Nice: non-linear independent components estimation. arXiv preprint, arXiv:1410.8516 (2014). [9] F. Noe and H. Wu, Boltzmann Generators – Sampling Equilibrium States of Many- Body Systems with Deep Learning. arXiv preprint, arXiv:1812.01729 (2018). [10] M. S. Albergo, G. Kanwar, and P. E. Shanahan, Flow-based generative models for Markov chain Monte Carlo in lattice field theory. arXiv preprint, arXiv: 1904.12072v2 (2019).