Google, UCLA prompt AI for better answer choices

Google, UCLA prompt AI for better answer choices

AI Breakthrough: Google’s AVIS Program Takes a Dynamic Approach to Answering Questions

Google’s AVIS program

Artificial intelligence (AI) has continuously amazed the public with its ability to generate answers from any query. However, the accuracy of these answers has sometimes fallen short, as programs like ChatGPT lack subject-specific knowledge and can even produce false information. But now, researchers at the University of California and Google have developed a new approach, called AVIS (Autonomous Visual Information Seeking with Large Language Models), that allows AI programs to dynamically select the most appropriate tool, such as web search or optical character recognition, to find answers from alternative sources. The result is a significant step towards achieving “planning” and “reasoning” capabilities in AI.

The AVIS program builds on Google’s Pathways Language Model (PaLM), a large language model that has been used in various generative AI experiments. It is part of a broader trend in AI research that aims to turn machine learning programs into more versatile agents, capable of complex tasks beyond text prediction. Examples of such endeavors include BabyAGI, an AI-powered task management system, and PaLM*E, which enables a robot to execute a series of physical actions.

What sets AVIS apart is its ability to adapt its actions on the fly, rather than following a pre-determined course of action like BabyAGI and PaLM*E. AVIS employs an algorithm called a “Planner” to evaluate prompted text, break it down into sub-questions, and generate a set of possible actions. It even incorporates human decision-making by using a “transition graph,” which models how humans choose tools in different situations. This integration of human choices adds a human-like touch to AVIS’s decision-making process.

To ensure the accuracy of its choices, AVIS is equipped with a “Reasoner” algorithm, which evaluates the usefulness of each chosen tool. If a tool proves ineffective, the Reasoner sends the Planner back to explore alternative options. This iterative process ensures that AVIS produces satisfactory answers to the original questions.

The performance of AVIS was tested on standard benchmarks for visual question answering, such as OK-VQA. The results were impressive, with AVIS achieving an accuracy of 60.2%, surpassing many existing methods tailored specifically for this dataset. This demonstrates the potential of AVIS and its dynamic decision-making framework to outperform task-specific AI models, emphasizing the increasing generality of machine learning AI.

Going forward, the researchers behind AVIS have plans to extend its capabilities beyond image-based questions. They envision a future where AVIS can tackle a wide range of reasoning tasks, further solidifying its position as a groundbreaking AI breakthrough.

AVIS workflow

In conclusion, Google’s AVIS program represents a significant step forward in AI technology. By incorporating dynamic decision-making and learning from human choices, AVIS offers a more refined and accurate approach to question answering. With its compelling performance on benchmark tests, AVIS showcases the potential for machine learning AI to excel in a broad range of tasks. As AVIS continues to evolve, we can expect even more astonishing applications of AI in the near future.

References:ZDNet ArticleResearch Paper on arXiv