Learning (Approximately) Equivariant Networks via Constrained Optimization

We are officially relaunching the Bucharest Deep Learning meetup! Join us for a deep dive into equivariant neural networks with Andrei Manolache, presenting his recent NeurIPS 2025 Oral paper.
The Talk: Learning (Approximately) Equivariant Networks via Constrained Optimization Strictly equivariant neural networks are great for perfect data, but real-world data is noisy and breaks these symmetries. On the flip side, fully unconstrained models miss out on structural advantages.
The paper introduces ACE (Adaptive Constrained Equivariance): a novel optimization approach that starts with a flexible model and gradually enforces equivariance. This method finds the perfect data-driven balance, significantly improving sample efficiency and robustness compared to strict models.

Logistics:

• Date & Time: Thursday, February 26 | 18:30 - 19:30
• Location: FMI New Building (Politehnica Business Tower)
• Address: Bulevardul Iuliu Maniu, nr. 15G, Etaj 5, Room 503