Presentation at NeurIPS 2022, “Symmetries and Geometry in Neural Representations” Workshop (NeurReps)
Workshop website | NeurIPS website
Last week, some of our members went to New Orleans for the thirty-sixth Conference on Neural Information Processing Systems, one of the biggest gatherings of machine learning researchers from all over the world. Having the chance to bring part of our association to NeurIPS 2022 was a great honour for us.
We took part in multiple events. Simone Azeglio and Arianna Di Bernardo co-organised with Sophia Sanborn, Nina Miolane and Christian Shewmake the wonderful Symmetry and Geometry in Neural Representations (NeurReps) workshop gathering together researchers at the intersection of geometric deep learning, applied geometry, and neuroscience (Taco Cohen, Irina Higgins and many more) to study the geometric principles underlying neural representations.
Secondly, our members Luca Bottero, Francesco Calisto and Valerio Pagliarino presented their work on learning geometrical features from images through group action enforcement in the poster session (https://openreview.net/forum?id=sEn61s0M1hy).
They proposed an autoencoder architecture capable of automatically learning meaningful geometric features of objects in images, achieving a disentangled representation of 2D objects. It is made of a standard dense autoencoder that captures the deep features identifying the shapes and an additional encoder that extracts geometric latent variables regressed in an unsupervised manner.
Simone Azeglio also presented his work as part of the SENSORIUM competition on predicting large scale mouse primary visual cortex activity.
Conferences like NeurIPS are a unique stage where to find inspiring insights in other people’s work and to create connections in both the academic and industrial domains. That is why we are grateful to TesiSquare, Fondazione DIG421 and NeurIPS for their great support and to HPC4AI, NPO Sistemi for the help. Last but not least, we thank our institutions LMU, TUM, École Normale Supérieure and UniTO.
Abstract:
https://openreview.net/forum?id=sEn61s0M1hy
Code:
Poster:
