Google unveiled Minerva AI, a sophisticated natural language processing model that can handle complex computations and solve math questions. It is a neural network dubbed PaLM that came into existence in April, 2022. PaLM features around 540 billion parameters that determine how this particular AI model makes decisions. Based on the parameters, Google’s Minerva has also surpassed OpenAI’s GPT-3, with 175 billion parameters.
Google claims that existing neural networks have only so far shown a modest capacity to handle so-called quantitative reasoning issues, such as math problems. Ethan Dyer and Guy Gur-Ari, Google researchers, wrote that most language models tend to overlook quantitative reasoning and still fall short of human-grade performance. They added, “It is often believed that solving quantitative reasoning problems using machine learning will require significant advancements in model architecture and training techniques.”
Read More: Calantic AI Marketplace by Bayer analyzes CT and MRI Scans.
118 GB of scientific papers and web pages containing mathematical expressions served as the dataset used by Google to train Minerva. The neural network retained the formatting and symbols that express the semantic meaning of mathematical expressions. This is why Minerva learned to express answers using standard mathematical notation.
Additionally, Google set up Minerva to produce a range of potential responses when analyzing a query. Minerva can even come up with various ways to calculate the exact answer regarding math difficulties. After all the computations, the neural network compares the solutions obtained from different calculations and choose the most appropriate one.
Dyer and Gur-Ari concluded, ”We hope that general models capable of solving quantitative reasoning problems will help push the frontiers of science and education.”
