Research Program 3 (R3)

Next-generation data-efficient deep learning

FCAI Research Program Next-generation data-efficient deep learning contributes to research objective Data efficiency (objective 1) by developing methods which harness the power of deep learning while achieving good results with less training data and in particular less human supervision. These methods include semi-supervised learning, few-shot learning for making use of auxiliary sources of training data, and learning models that can be reliably used in simulator-based inference.

The research in Next-generation data-efficient deep learning aims to create one of the first AI systems capable of limited analogue reasoning. New insights will be brought by using simulation techniques to deep learning techniques, such as generative adversarial learning (GAN) and variational autoencoders (VAE) through their connections to Approximate Bayesian Computation (ABC). This will enable applying deep neural networks in industrial process optimization, enhanced perception in autonomous driving, or novel human-computer interaction.

Examples of publications:

• Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, David Lopez-Paz. 2022. Interpolation consistency training for semi-supervised learning. Neural Networks, 145:90–106.

• Subhankar Roy, Martin Trapp, Andrea Pilzer, Juho Kannala, Nicu Sebe, Elisa Ricci, Arno Solin. 2022. Uncertainty-guided source-free domain adaptation. Proceedings of European Conference on Computer Vision (ECCV).

• I-Ju Chen, Markus Aapro, Abraham Kipnis, Alexander Ilin, Peter Liljeroth, Adan S. Foster. 2022. Precise atom manipulation through deep reinforcement learning. Nature Communications, 13, 7499.

• Ari Heljakka, Yuxin Hou, Juho Kannala, Arno Solin. 2020. Deep Automodulators. Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS), Advances in Neural Information Processing Systems, 33.

• Lassi Meronen, Christabella Irwanto, Arno Solin. 2020. Stationary Activations for Uncertainty Calibration in Deep Learning. Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS), Advances in Neural Information Processing Systems, 33.

Program poster (pdf)

Coordinating professor: Arno Solin

People

The groups of following PIs take part in the Research Program Next-generation data-efficient deep learning. If you would like to join this program, please contact the coordinating professor.

• Luigi Acerbi, University of Helsinki

• Vikas Garg, Aalto University

• Stéphane Deny, Aalto University

• Aapo Hyvärinen, University of Helsinki

• Perttu Hämäläinen, Aalto University

• Juho Kannala, Aalto University

• Samuel Kaski, Aalto University

• Arto Klami, University of Helsinki

• Mikko Kurimo, Aalto University

• Leo Kärkkäinen, Aalto University

• Jorma Laaksonen, Aalto University

• Jaakko Lehtinen, Aalto University, NVIDIA

• Pekka Marttinen, Aalto University

• Antti Oulasvirta, Aalto University

• Patrick Rinke, Aalto University

• Arno Solin, Aalto University – coordinating professor

• Jörg Tiedemann, University of Helsinki

• Harri Valpola, Curious AI

• Aki Vehtari, Aalto University

• Roman Yangarber, University of Helsinki

• Bo Zhao, Aalto University