Dreaming More Data: Class-dependent Distributions over Diffeomorphisms for Learned Data Augmentation

Søren Hauberg, Oren Freifeld, Anders Boesen Lindbo Larsen, John Fisher, Lars Hansen
; Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, PMLR 51:342-350, 2016.

Abstract

Data augmentation is a key element in training high-dimensional models. In this approach, one synthesizes new observations by applying pre-specified transformations to the original training data; e.g. new images are formed by rotating old ones. Current augmentation schemes, however, rely on manual specification of the applied transformations, making data augmentation an implicit form of feature engineering. With an eye towards true end-to-end learning, we suggest learning the applied transformations on a per-class basis. Particularly, we align image pairs within each class under the assumption that the spatial transformation between images belongs to a large class of diffeomorphisms. We then learn a class-specific probabilistic generative models of the transformations in a Riemannian submanifold of the Lie group of diffeomorphisms. We demonstrate significant performance improvements in training deep neural nets over manually-specified augmentation schemes. Our code and augmented datasets are available online.

Cite this Paper


BibTeX
@InProceedings{pmlr-v51-hauberg16, title = {Dreaming More Data: Class-dependent Distributions over Diffeomorphisms for Learned Data Augmentation}, author = {Søren Hauberg and Oren Freifeld and Anders Boesen Lindbo Larsen and John Fisher and Lars Hansen}, booktitle = {Proceedings of the 19th International Conference on Artificial Intelligence and Statistics}, pages = {342--350}, year = {2016}, editor = {Arthur Gretton and Christian C. Robert}, volume = {51}, series = {Proceedings of Machine Learning Research}, address = {Cadiz, Spain}, month = {09--11 May}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v51/hauberg16.pdf}, url = {http://proceedings.mlr.press/v51/hauberg16.html}, abstract = {Data augmentation is a key element in training high-dimensional models. In this approach, one synthesizes new observations by applying pre-specified transformations to the original training data; e.g. new images are formed by rotating old ones. Current augmentation schemes, however, rely on manual specification of the applied transformations, making data augmentation an implicit form of feature engineering. With an eye towards true end-to-end learning, we suggest learning the applied transformations on a per-class basis. Particularly, we align image pairs within each class under the assumption that the spatial transformation between images belongs to a large class of diffeomorphisms. We then learn a class-specific probabilistic generative models of the transformations in a Riemannian submanifold of the Lie group of diffeomorphisms. We demonstrate significant performance improvements in training deep neural nets over manually-specified augmentation schemes. Our code and augmented datasets are available online.} }
Endnote
%0 Conference Paper %T Dreaming More Data: Class-dependent Distributions over Diffeomorphisms for Learned Data Augmentation %A Søren Hauberg %A Oren Freifeld %A Anders Boesen Lindbo Larsen %A John Fisher %A Lars Hansen %B Proceedings of the 19th International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2016 %E Arthur Gretton %E Christian C. Robert %F pmlr-v51-hauberg16 %I PMLR %J Proceedings of Machine Learning Research %P 342--350 %U http://proceedings.mlr.press %V 51 %W PMLR %X Data augmentation is a key element in training high-dimensional models. In this approach, one synthesizes new observations by applying pre-specified transformations to the original training data; e.g. new images are formed by rotating old ones. Current augmentation schemes, however, rely on manual specification of the applied transformations, making data augmentation an implicit form of feature engineering. With an eye towards true end-to-end learning, we suggest learning the applied transformations on a per-class basis. Particularly, we align image pairs within each class under the assumption that the spatial transformation between images belongs to a large class of diffeomorphisms. We then learn a class-specific probabilistic generative models of the transformations in a Riemannian submanifold of the Lie group of diffeomorphisms. We demonstrate significant performance improvements in training deep neural nets over manually-specified augmentation schemes. Our code and augmented datasets are available online.
RIS
TY - CPAPER TI - Dreaming More Data: Class-dependent Distributions over Diffeomorphisms for Learned Data Augmentation AU - Søren Hauberg AU - Oren Freifeld AU - Anders Boesen Lindbo Larsen AU - John Fisher AU - Lars Hansen BT - Proceedings of the 19th International Conference on Artificial Intelligence and Statistics PY - 2016/05/02 DA - 2016/05/02 ED - Arthur Gretton ED - Christian C. Robert ID - pmlr-v51-hauberg16 PB - PMLR SP - 342 DP - PMLR EP - 350 L1 - http://proceedings.mlr.press/v51/hauberg16.pdf UR - http://proceedings.mlr.press/v51/hauberg16.html AB - Data augmentation is a key element in training high-dimensional models. In this approach, one synthesizes new observations by applying pre-specified transformations to the original training data; e.g. new images are formed by rotating old ones. Current augmentation schemes, however, rely on manual specification of the applied transformations, making data augmentation an implicit form of feature engineering. With an eye towards true end-to-end learning, we suggest learning the applied transformations on a per-class basis. Particularly, we align image pairs within each class under the assumption that the spatial transformation between images belongs to a large class of diffeomorphisms. We then learn a class-specific probabilistic generative models of the transformations in a Riemannian submanifold of the Lie group of diffeomorphisms. We demonstrate significant performance improvements in training deep neural nets over manually-specified augmentation schemes. Our code and augmented datasets are available online. ER -
APA
Hauberg, S., Freifeld, O., Larsen, A.B.L., Fisher, J. & Hansen, L.. (2016). Dreaming More Data: Class-dependent Distributions over Diffeomorphisms for Learned Data Augmentation. Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, in PMLR 51:342-350

Related Material