Deep Generative Models Strike Back! Improving Understanding and Evaluation in Light of Unmet Expectations for OoD Data
This addresses a critical problem in unsupervised learning for researchers and practitioners by improving OoD detection in generative models, though it is incremental as it builds on existing methods and datasets.
The paper tackles the issue of deep generative models assigning higher likelihood to out-of-distribution (OoD) data than in-sample data, showing that datasets like MNIST and CIFAR10 are trivially separable and that models like masked autoregressive flows avoid this problem better than others, while a change-of-basis method improves discrimination results and probabilistic PCA outperforms complex models in anomaly detection with less complexity and faster training.
Advances in deep generative and density models have shown impressive capacity to model complex probability density functions in lower-dimensional space. Also, applying such models to high-dimensional image data to model the PDF has shown poor generalization, with out-of-distribution data being assigned equal or higher likelihood than in-sample data. Methods to deal with this have been proposed that deviate from a fully unsupervised approach, requiring large ensembles or additional knowledge about the data, not commonly available in the real-world. In this work, the previously offered reasoning behind these issues is challenged empirically, and it is shown that data-sets such as MNIST fashion/digits and CIFAR10/SVHN are trivially separable and have no overlap on their respective data manifolds that explains the higher OoD likelihood. Models like masked autoregressive flows and block neural autoregressive flows are shown to not suffer from OoD likelihood issues to the extent of GLOW, PixelCNN++, and real NVP. A new avenue is also explored which involves a change of basis to a new space of the same dimension with an orthonormal unitary basis of eigenvectors before modeling. In the test data-sets and models, this aids in pushing down the relative likelihood of the contrastive OoD data set and improve discrimination results. The significance of the density of the original space is maintained, while invertibility remains tractable. Finally, a look to the previous generation of generative models in the form of probabilistic principal component analysis is inspired, and revisited for the same data-sets and shown to work really well for discriminating anomalies based on likelihood in a fully unsupervised fashion compared with pixelCNN++, GLOW, and real NVP with less complexity and faster training. Also, dimensionality reduction using PCA is shown to improve anomaly detection in generative models.