We investigate the construction of generative models capable of encoding physical constraints that can be hard to express explicitly. For the problem of inverse material design, where one seeks to design a material with a prescribed set of properties, a significant challenge is ensuring synthetic viability of a proposed new material. We encode an implicit dataset relationships, namely that certain materials can be decomposed into other ones in the dataset, and present a VAE model capable of preserving this property in the latent space and generating new samples with the same. This is particularly useful in sequential inverse material design, an emergent research area that seeks to design a material with specific properties by sequentially adding (or removing) elements using policies trained through deep reinforcement learning.
Demystifying Inductive Biases for $β$-VAE Based Architectures
Max Planck Institute for Intelligent Systems(马克斯·普朗克智能系统研究所)
$β$ -变分自动编码器($β$ -VAE)及其变体在学习语义上有意义的解纠缠表示方面的性能是无与伦比的。另一方面,有理论论据表明,无监督的解缠是不可能的。在这项工作中,我们揭示了负责基于VAE的架构的成功的感性偏见。我们表明,在经典的数据集的方差结构,诱导的生成因子,是方便地对齐的潜在方向培育的VAE目标。这就建立了VAE的解开能力所依赖的关键偏差。通过对现有数据集进行小而精细的扰动,我们隐藏了方便的相关结构,这种结构很容易被各种架构所利用。为了证明这一点,我们构建了标准数据集的修改版本,其中(i)生成因子被完美地保留;(ii)每个图像都经过温和的变换,导致方差的微小变化;(iii)领先的\textbf{VAE-based disentanglement architectures fail to produce disentangled representations whilst the performance of a non-variational method remains unchanged}。我们的修改的建设是不平凡的,并依赖于最近的进展机制的理解$β$ -VAE及其连接到PCA。我们通过提供具有独立利益的额外见解来加强这种联系。