Skip to main content
Vai all'homepage della Commissione europea (si apre in una nuova finestra)
italiano it
CORDIS - Risultati della ricerca dell’UE
CORDIS

Geometric Deep Learning for Neural Artifacts: Symmetry-Aware Learning across Trained Model Weights, Internal Representations, and Gradients

Obiettivo

The explosive growth of deep learning has created fundamentally new data modalities: the mathematical byproducts of neural network development. Practitioners generate daily vast amounts of valuable but underutilized data---trained model weights, internal representations from inference, and network gradients from training. These byproducts, which we term Neural Artifacts, contain crucial information about model behavior, optimization, and internal processing.
Learning on neural artifacts holds transformative potential. Learning on Weight Spaces - weights of trained models - could revolutionize model selection, editing and generation. Learning on internal representations enables efficient hallucination detection on Large Language Models and other analyses. Learning on gradient spaces could transform model optimization and analysis. However, existing methods for learning on neural artifacts achieve only a fraction of this potential as they fail to account for fundamental mathematical structure, leaving the field without principled methods to harness the full power of these emerging data types.
The key insight is that neural artifacts share deep mathematical structures—symmetries inherited from their underlying architecture. By developing a geometric deep learning framework that respects these symmetries, we can unlock their full potential. Building on our pioneering equivariant architectures for neural artifacts, we propose developing theoretical foundations and practical tools. This includes designing symmetry-aware architectures, establishing theoretical guarantees, developing novel applications, and delivering open-source implementations.
Success will provide practitioners with powerful tools for understanding, generating, optimizing, and controlling neural artifacts. This research transforms neural computation byproducts into learnable information— extremely challenging due to their high dimensionality and symmetry structure, but with revolutionary potential.

Campo scientifico (EuroSciVoc)

Questo progetto non è ancora stato classificato con EuroSciVoc.
Suggerisci i campi scientifici che ritieni più rilevanti e aiutaci a migliorare il nostro servizio di classificazione.

Parole chiave

Programma(i)

Argomento(i)

Meccanismo di finanziamento

HORIZON-ERC - HORIZON ERC Grants

Vedi tutti i progetti finanziati nell’ambito di questo schema di finanziamento

Invito a presentare proposte

(si apre in una nuova finestra) ERC-2026-STG

Vedi tutti i progetti finanziati nell’ambito del bando

Istituzione ospitante

TECHNION - ISRAEL INSTITUTE OF TECHNOLOGY
Contributo netto dell'UE
€ 1 637 500,00
Costo totale
€ 1 637 500,00

Beneficiari (1)