Call for papers

Deep learning has shown impressive performance improvements on a variety of tasks in vision, speech and language domains. However, a large amount of labeled data is often needed to deliver these remarkable accuracy gains. The goal of this workshop is to facilitate discussion on the role of equivariance/ invariance of feature maps, including their synergies with unsupervised/self-supervised learning methods that define and solve auxiliary tasks on unlabeled data to learn representations (such as auto-encoding, context prediction, predicting one modality from other, etc.), towards reducing the dependence on labeled examples. Apart from the group-theoretical notions of equivariance and invariance, the workshop also welcomes contributions with a non-group-theoretical and relaxed notion of equivariance and invariance with respect to a set of transformations (possibly input specific and learned from data).

We welcome submissions on topics including (but not limited to):

  • Learning desired equivariances / invariances for a given prediction task from data (e.g., from massive amounts of unlabeled data, from auxiliary labels, from multiple views, or from side-information when available)
  • Learning transferable invariances: feature maps that are equivariant/invariant with respect to domain- or task-specific nuisance factors while being suitable for transfer learning (on a target domain or task)
  • Learning disentangled representations for label efficiency 
  • Learning equivariant / invariant feature maps from structure when it is available (e.g., temporal ordering structure in videos)
  • Incorporating known equivariances / invariances (from domain knowledge, e.g., rotation and scale for images, permutations for sets, etc.) as inductive biases
  • Architectural priors for equivariant features 
  • Synergies with unsupervised / self-supervised feature learning methods that define and solve auxiliary tasks on unlabeled data (such as auto-encoding, context prediction, etc.) to learn representation maps
  • Hierarchical representations with interplay between invariance and equivariance 
  • Visualizing/understanding the equivariances and invariances learned by current popular deep neural net architectures

Submission instructions.  

Submission website:
Submissions should be a maximum of 4 pages (excluding references, using the ICML format) and should not have been published/presented at other venues. However, work currently under review but not accepted anywhere is welcome. The submissions can contain author details. The accepted papers might be posted on the workshop website but the workshop will not have archival proceedings. All accepted papers will be presented at the poster sessions. Selected top submissions will also be given spotlight presentation slots. 


[1] Fabio Anselmi, Georgios Evangelopoulos, Lorenzo Rosasco, and Tomaso Poggio. Symmetry regularization. Technical report, Center for Brains, Minds and Machines (CBMM), 2017. 

[2] Taco Cohen and Max Welling. Group equivariant convolutional networks. In Proceedings of the 33rd International Conference on Machine Learning, 2016.

[3] Taco S Cohen and Max Welling. Steerable cnns. In ICLR, 2017.

[4] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision, pages 1422–1430, 2015.

[5] William T Freeman, Edward H Adelson, et al. The design and use of steerable filters. IEEE Transactions on Pattern analysis and machine intelligence, 13(9):891–906, 1991.

[6] Geoffrey E Hinton, Sara Sabour, Nicholas Frosst. Matrix capsules with EM routing. ICLR, 2018.

[7] Jörn-Henrik Jacobsen, Bert de Brabandere, and Arnold WM Smeulders. Dynamic steerable blocks in deep residual networks. arXiv preprint arXiv:1706.00598, 2017.

[8] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. In Advances in Neural Information Processing Systems, pages 2017–2025, 2015.

[9] Dinesh Jayaraman and Kristen Grauman. Learning image representations equivariant to ego-motion. In Proc. ICCV, 2015.

[10] Yannic Kilcher, Gary Becigneul, and Thomas Hofmann. Parametrizing filters of a cnn with a gan. arXiv preprint arXiv:1710.11386, 2017.

[11] Abhishek Kumar, Prasanna Sattigeri, and Tom Fletcher. Semi-supervised learning with gans: Manifold invariance with improved inference. In Advances in Neural Information Processing Systems, pages 5540–5550, 2017.

[12] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 991–999, 2015.

[13] Youssef Mroueh, Stephen Voinea, and Tomaso A Poggio. Learning with group invariant features: A kernel perspective. In Advances in Neural Information Processing Systems, pages 1558–1566, 2015.

[14] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2536–2544, 2016.

[15] Anant Raj, Abhishek Kumar, Youssef Mroueh, P Thomas Fletcher, and Bernhard Schölkopf. Local group invariant representations via orbit embeddings. In AISTATS, 2017.

[16] Siamak Ravanbakhsh, Jeff Schneider, and Barnabas Poczos. Equivariance through parameter-sharing. In ICML, 2017.

[17] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. arXiv preprint arXiv:1711.07289, 2017.