Google has announced the launch of a new framework for artificial intelligence called “Social Learning” that aims to improve the collaborative workability of intelligent language models without negatively impacting individuals’ privacy.
This system allows artificial intelligence systems to engage in mutual learning through natural language communication, leading to information sharing and increased efficiency in performing various tasks.
Using Google’s new approach, models used by teachers can share knowledge with models used by students without the need to directly share sensitive or personal data, thus preserving privacy while facilitating the learning process.
In this mechanism, students’ models draw their knowledge from a diverse set of teacher models, each mastering specific skills such as recognizing unwanted emails, solving mathematical problems, or providing answers to questions based on the given text.
By utilizing examples classified by humans, teacher models can train student models without the need to share raw data, addressing privacy issues arising from data sharing.
Furthermore, teaching models can create new examples or provide task-specific guidelines, significantly contributing to enhancing the learning process.
Experiments have shown that the “Social Learning” method has been effective in improving students’ performance in a variety of tasks.
The optimized models created by experts have elevated efficiency to levels comparable to real-world data while significantly reducing privacy violations.
The guidelines generated by teacher models have proven effective in enhancing the performance of student models, indicating the vast potential of language models in adapting to given commands.
To ensure privacy preservation, researchers have utilized tools like “Secret Sharer” to assess information leakage during the learning process.
Evidence suggests a very low rate of information leakage, enhancing the credibility of the educational framework in carrying out tasks without disclosing private information from the core data set.
By employing mechanisms resembling human social learning, these systems can efficiently transfer information and enhance each other’s capabilities while guaranteeing user privacy protection.
This path represents progress in developing artificial intelligence systems that respect privacy across various sectors.
Researchers seek to enhance the structure of the “Social Learning” system and explore its diverse applications in tasks and various types of data.
