Initiatives for training machine learning models with open data

Behind a voice-enabled virtual assistant, a movie recommendation on a streaming platform, or the development of some COVID-19 vaccines, there are machine learning models. This branch of artificial intelligence enables systems to learn and improve their performance.

Machine learning (ML) is one of the fields driving technological progress today, and its applications are growing every day. Examples of solutions developed with machine learning include DALL-E, the set of language models in Spanish known as MarIA, and even Chat GPT-3, a generative AI tool capable of creating content of all types, such as code for programming data visualizations from the datos.gob.es catalog.

All of these solutions work thanks to large data repositories that make system learning possible. Among these, open data plays a fundamental role in the development of artificial intelligence as it can be used to train machine learning models.

Based on this premise, along with the ongoing effort of governments to open up data, there are non-governmental organizations and associations that contribute by developing applications that use machine learning techniques aimed at improving the lives of citizens. We highlight three of them:

ML Commons is driving a better machine learning system for everyone

This initiative aims to improve the positive impact of machine learning on society and accelerate innovation by offering tools such as open datasets, best practices, and algorithms. Its founding members include companies such as Google, Microsoft, DELL, Intel AI, Facebook AI, among others.

According to ML Commons, around 80% of research in the field of machine learning is based on open data. Therefore, open data is vital to accelerate innovation in this field. However, nowadays, "most public data files available are small, static, legally restricted, and not redistributable," as David Kanter, director of ML Commons, assures.

In this regard, innovative ML technologies require large datasets with licenses that allow their reuse, that can be redistributed, and that are continually improving. Therefore, ML Commons' mission is to help mitigate that gap and thus promote innovation in machine learning.

The main goal of this organization is to create a community of open data for the development of machine learning applications. Its strategy is based on three pillars:

Firstly, creating and maintaining comprehensive open datasets, including The People's Speech, with over 30,000 hours of speech in English to train natural language processing (NLP) models, Multilingual Spoken Words, with over 23 million expressions in 50 different languages, or Dollar Street, with over 38,000 images of homes from around the world in various socio-economic situations. The second pillar involves promoting best practices that facilitate standardization, such as the MLCube project, which proposes standardizing the container process for ML models to facilitate shared use. Lastly, benchmarking in study groups to define benchmarks for the developer and research community.

Taking advantage of the benefits and being part of the ML Commons community is free for academic institutions and small companies (less than ten workers).

Datacommons synthesizes different sources of open data into a single portal

Datacommons aims to enhance democratic data flows within the cooperative and solidarity economy and its main objective is to offer purified, normalized, and interoperable data.

The variety of formats and information offered by public portals of open data can be a hindrance to research. The goal of Datacommons is to compile open data into an encyclopedic website that organizes all datasets through nodes. This way, users can access the source that interests them the most.

This platform, designed for educational and journalistic research purposes, functions as a reference tool for navigating through different sources of data. The team of collaborators works to keep the information up-to-date and interacts with the community through its email (support@datacommons.org) or GitHub forum.

Papers with Code: the open repository of materials to feed machine learning models

This is a portal that offers code, reports, data, methods, and evaluation tables in open and free format. All content on the website is licensed under CC-BY-SA, meaning it allows copying, distributing, displaying, and modifying the work, even for commercial purposes, by sharing the contributions made with the same original license.

Any user can contribute by providing content and even participate in the community's Slack channel, which is moderated by responsible individuals who protect the platform's defined inclusion policy.

As of today, Papers with Code hosts 7806 datasets that can be filtered by format (graph, text, image, tabular, etc.), task (object detection, queries, image classification, etc.), or language. The team maintaining Papers with Code belongs to the Meta Research Institute.

The goal of ML Commons, Data Commons, and Papers with Code is to maintain and grow open data communities that contribute to the development of innovative technologies, including artificial intelligence (machine learning, deep learning, etc.) with all the possibilities its development can offer to society.

As part of this process, the three organizations play a fundamental role: they offer standard and redistributable data repositories to train machine learning models. These are useful resources for academic exercises, promoting research, and ultimately facilitating the innovation of technologies that are increasingly present in our society.