GitHub - Open-Weights/Definition: Definition for Open Weights LIcensing

The article addresses the fundamental misconception that open source software licensing principles can be directly applied to Neural Network Weights (NNWs). It highlights that NNWs are mathematical matrices representing learned knowledge, unlike human-readable source code. Because weights cannot be easily studied, debugged, or modified by humans, the traditional freedoms of running, studying, and altering software do not translate effectively to this new artifact. This distinction necessitates a reevaluation of how openness is defined in the AI era. To resolve this, the text argues for the urgent development of standardized "Open Weights" licensing frameworks tailored specifically to the unique nature of NNWs. Rather than relying on government regulation, which has failed to produce consistent standards, the initiative promotes voluntary community-driven definitions aligned with the core ideals of free software. This approach ensures that legal and practical guidelines for sharing and using model weights are established by developers and experts, preserving the original goals of openness without importing complex issues like data privacy or human rights into the license terms. Relevance to open data lies in its pioneering effort to adapt transparency and sharing norms to machine learning artifacts. By creating a public, community-vetted definition and permissive license for weights, it establishes a precedent for making AI models accessible and verifiable. This fosters greater trust, reproducibility, and collaboration within the AI ecosystem, ensuring that the benefits of open methodologies extend beyond traditional code to the foundational knowledge encoded in neural networks.

Source: github.com
Published on 2023-07-10

Defiende INAI su trabajo en favor de la transparencia »

Open Data

Related news