Overview

Date	One-day workshop on Sunday April 27th or Monday April 28th, 2025
Location	The workshop will be held in person in Singapore. Remote access will be made available at a later date.

The recent surge in the number of publicly available neural network models—exceeding a million on platforms like Hugging Face—calls for a shift in how we perceive neural network weights. This workshop aims to establish neural network weights as a new data modality, offering immense potential across various fields.

We plan to address key dimensions of weight space learning:

• Weight Space as a Modality: Characterization of the weight space,
  symmetries, scaling laws, and model zoo datasets.
• Model analysis: Inferring model properties, such as test performance
  or generalization error, by the inspection of their weights.
• Model synthesis: Generating model weights for specific datasets and tasks,
  and improving tasks like pruning, merging, and robustness through weight manipulation.
• Learning from populations: Using data from neural network populations to identify
  and exploit structure in trained models.
• Applications to neural fields and beyond: Applying weight space learning to 3D shape analysis, neural radiance fields (NeRFs), and other domains.

Weight space learning remains a nascent and scattered research area. Our goal is to provide a bridge between the abovementioned topics, and research areas such as model merging, neural architecture search, and meta-learning. By aligning terminology and methodologies, we aim to drive sustained progress and foster interdisciplinary collaboration.

Research Goals and Key Questions

This workshop will explore fundamental questions about weight spaces, such as:

• What properties of weights, such as symmetries and invariances, present challenges or can be leveraged for optimization, learning and generalization?
• How can model weights be efficiently represented, manipulated, and used for downstream tasks?
• What model information can be decoded from model weights?
• Can model weights be generated for specific applications, to make training and model selection more efficient?
• Can weight space learning benefit research in processing and synthesising neural fields, for e.g. scientific applications and 3D vision?
• How can we democratize the usage of weight spaces, enabling more efficient research progress?

Accessibility and Modality

All accepted papers and posters will be available online for remote attendees. The talks will be recorded and made available on YouTube. We will also have a dedicated Discord channel for virtual engagement, ensuring that participants, whether in-person or remote, can collaborate and discuss ideas.