Last News
Artificial Intelligence (AI) has become a fundamental part of our lives, but its complexity can be overwhelming. In this sense, the PAIR (People + AI Research) team at Google Research is leading a revolutionary initiative to make AI more accessible and understandable for everyone. By implementing advances in Generative AI, visualization and educational tools, transparency, and software development, PAIR is changing the way we approach AI and opening up new opportunities for the technology research community.
they comment in your article that its mission is to build AI systems with people in mind from the beginning of the process.
Generative AI: Expanding the boundaries of creativity
The field of Genetic AI has attracted the attention of the scientific community and has fueled great excitement about the creative possibilities of AI. PAIR is immersed in a series of investigations related to this topic, from the use of language models to create generative agents to the study of how artists adopt generative models of images.
Text-to-image models allow people to enter a text-based description to generate a corresponding image. For example, you can write “a gingerbread house in a forest in cartoon style.” In a recent study, entitled “The Pras Artists” and to be published in Creativity and Cognition 2023, PAIR discovered that the wearers of these models are not only trying to create beautiful images, but also to develop unique and innovative styles. To achieve these styles, some even search domain-specific vocabulary, such as consulting architecture blogs, to produce specific images of buildings.
PAIR is also researching solutions to challenges faced by instructional creators working with generative AI. Basically these creators do programming without using a normal programming language. To address this challenge, PAIR has developed new methods to extract semantic structures from natural language instructions. These structures were implemented in statement editors to provide features similar to those found in other programming environments, such as semantic highlighting, autosuggestions, and comments on structured data.
Agile Classifiers: Towards safer and more moderated online speech
Moderation of online content and the promotion of safe speech is a critical aspect of interacting with AI. PAIR developed so-called “Agile Classifiers” as a solution to address this challenge. These classifiers take advantage of the semantic and syntactic strengths of LLMs to solve classification problems related to toxic speech online.
One of the main advantages of agile classifiers is their ability to be developed with very small data sets, even with as few as 80 samples. This means that security classifiers tailored to specific use cases can be created in a short period of time, making it easy to quickly adapt and correct unwanted biases in models. These methods recently won a SemEval competition to identify and explain sexism online, proving their usefulness in improving content moderation.
Agile classifiers have also enabled the development of new explanatory methods to identify the role of training data in behavior and potential model errors. By combining these training data allocation methods with agile classifiers, it was possible to identify mislabeled samples in the training data sets. This allows to reduce the noise in the training data and significantly improve the accuracy of the generated models.
Visualization and educational tools
To make the field of AI more accessible and understandable, PAIR is committed to designing and publishing highly visual and interactive online essays called “AI Explorables”. These audit materials provide a practical and accessible way to learn about key concepts in machine learning.
One of the most recently published review articles, entitled “From Confidently Incorrect Models to Humble Ensembles”, addresses the problem of model confidence. It explores why models can sometimes be very confident in their predictions and still be wrong. The explorer uses interactive examples to demonstrate how models with greater confidence in their predictions can be built using a technique called assimilation, which involves averaging the outputs of multiple models.
PAIR has also developed modern visualization methods to identify the role of training data in model behavior and errors. These tools allow you to identify behavioral problems in the models and provide better insight to improve the generation of results.
Transparency and Data Card Initiative
Transparency is a critical issue in the field of AI, and PAIR has worked to develop tools to address this challenge. In collaboration with the Technology, AI, Society and Culture (TASC) team, PAIR presented the “Data Cards” at the ACM FAccT’22 conference and launched the “Data Cards Playbook” as an open resource.
The Data Cards Playbook is a comprehensive guide that offers engagement tools and frameworks to help teams and organizations establish transparency practices. The playbook draws on the experience of more than 20 teams at Google and provides resources such as scalable frameworks, evidence-based guidance, and cross-disciplinary workshops to address team transparency challenges. It also includes an interactive lab that allows you to generate interactive “Data Cards” from text in mark-down format.
software tools
The PAIR team at Google Research is dedicated to developing a variety of software tools that improve understanding and access to AI models. These include Know Your Data, a tool that allows researchers to test a model’s performance in a variety of scenarios through interactive qualitative exploration of data sets. This tool helps identify and correct unwanted biases in training data.
PAIR recently released version 0.5 of the “Learning Interpretation Tool” (LIT), an open source platform for visualizing and understanding AI models. The new version of ITL provides support for image and table data, new interpreters for tabular feature assignment, a visualization called “Dive” for feature data exploration, and performance improvements that allow ITL to scale up to 100,000 dataset entries .
PAIR also helped develop “MakerSuite”, a tool for rapidly prototyping new AI models using instruction-based programming. This tool simplifies the process of prototyping AI models, allowing people with varying levels of experience to create prototypes in minutes instead of months.
Towards a more inclusive future for AI
The work being done by the PAIR team at Google Research is critical to changing the way we interact with AI. Making AI more accessible, understandable and transparent is key to making this technology a useful and beneficial tool for all.
As the field of AI advances, PAIR continues to develop new tools, research and educational materials to open up new possibilities and change the way people think about what they can achieve with AI. His commitment to transparency, education and accessibility is reflected in all of his projects.
Ultimately, PAIR’s goal is to build a future where AI is a tool that everyone can use and understand, opening doors to creativity, innovation and equal opportunity.