Talk by Zhuoqian Yang at University of St. Gallen

Weight Space Representation Learning through Neural Field Adaptation

The talk presented recent research on using neural network weights as meaningful data representations. This work differs from previous weight space learning approaches: rather than building model-agnostic weight encoders, the focus is on enforcing structure directly on the weight space, making the weights themselves serve as effective representations without additional encoding steps. By constraining the optimization space through a pre-trained base model and multiplicative low-rank adaptation (LoRA), neural field weights can be transformed into structured representations that capture semantic properties of the encoded data. The approach was validated across reconstruction, generation, and analysis tasks on 2D and 3D data, demonstrating that multiplicative LoRA weights enable higher-quality generation than existing weight-space methods.

Slides are available here.