Conference item icon

Conference item

Do deep generative models know what they don't know?

Abstract:
A neural network deployed in the wild may be asked to make predictions for inputs that were drawn from a different distribution than that of the training data. A plethora of work has demonstrated that it is easy to find or synthesize inputs for which a neural network is highly confident yet wrong. Generative models are widely viewed to be robust to such mistaken confidence as modeling the density of the input features can be used to detect novel, out-of-distribution inputs. In this paper we challenge this assumption. We find that the density learned by flow-based models, VAEs, and PixelCNNs cannot distinguish images of common objects such as dogs, trucks, and horses (i.e. CIFAR-10) from those of house numbers (i.e. SVHN), assigning a higher likelihood to the latter when the model is trained on the former. Moreover, we find evidence of this phenomenon when pairing several popular image data sets: FashionMNIST vs MNIST, CelebA vs SVHN, ImageNet vs CIFAR-10 / CIFAR-100 / SVHN. To investigate this curious behavior, we focus analysis on flow-based generative models in particular since they are trained and evaluated via the exact marginal likelihood. We find such behavior persists even when we restrict the flow models to constant-volume transformations. These transformations admit some theoretical analysis, and we show that the difference in likelihoods can be explained by the location and variances of the data and the model curvature. Our results caution against using the density estimates from deep generative models to identify inputs similar to the training distribution until their behavior for out-of-distribution inputs is better understood.
Publication status:
Published
Peer review status:
Peer reviewed

Actions

Access Document

Files:

Authors

More by this author
Institution:
University of Oxford
Division:
MPLS
Department:
Statistics
Oxford college:
University College
Role:
Author
ORCID:
0000-0001-5365-6933


Host title:
International Conference on Learning Representations
Journal:
International Conference on Learning Representations More from this journal
Publication date:
2019-05-06
Acceptance date:
2018-12-21
Event location:
New Orleans, USA


Pubs id:
pubs:936925
UUID:
uuid:1e34f829-8de6-4fc9-954d-089b0f4950f0
Local pid:
pubs:936925
Source identifiers:
936925
Deposit date:
2019-02-06
ARK identifier:

Terms of use


Views and Downloads

Views and downloads will return soon






If you are the owner of this record, you can report an update to it here: Report update to this record

TO TOP