Sharing visual categories through language

As children, we learn categories through visual examples, verbal explanations, or both, and are often guided by “teachers” — perhaps a parent or other adult. In contrast, academic research has primarily studied non-pedagogical learning where there is no active teacher, and learning based on visual examples, omitting verbal-based category learning. A recent paper in Cognition by Arseny Moskvichev and co-authors aims to close this gap.

Read More

Study: an overlooked weakness in a powerful machine learning tool

In the last two decades, researchers have reported success modeling high-dimensional chaotic behaviors with a simple but powerful machine-learning approach called reservoir computing. A new paper in Physical Review Research identifies limitations to reservoir computing and suggests a kind of catch-22 that can prove hard to circumvent, especially for complicated dynamic systems. 

Read More

Study: Visual Analogies for AI

The field of artificial intelligence has long been stymied by the lack of an answer to its most fundamental question: What is intelligence? To address questions about intelligence in AI, we need concrete tasks to pin down and test the notion of intelligence, argue SFI researchers Arseny Moskvichev, Melanie Mitchell, and Victor Vikram Odouard in a new paper in Transactions on Machine Learning Research.

Read More

Could AI ever truly "understand"?

ChatGPT knows how to use the word “tickle” in a sentence but it cannot feel the sensation. Can it then be said to understand the meaning of the word tickle the same way we humans do? In a paper for PNAS, SFI researchers Melanie Mitchell and David C. Krakauer survey the ongoing debate in which AI researchers are teasing apart whether Large Language Models like ChatGPT and Google’s PaLM understand language in any humanlike sense.

Read More

Research News Brief: Economics in nouns and verbs

In the last 50 years, economic theory has come to be based almost solely on mathematics. This brings logical precision, but according to a new paper by SFI economist Brian Arthur, it restricts what economics can easily talk about.

Read More

Study: Why are sustainable practices often elusive?

For at least 200,000 years, humans have been trying to understand their environments and adapt to them. At times, we have succeeded; often, we have not. In a new study, SFI's Stefani Crabtree, Jennifer Dunne, and others analyze how information flows from ecosystems to the societies inhabiting them.

Read More

Research News Brief: The frugal case for energy transition

If you think clean energy is expensive, try fossil fuels. A new report in the journal Joule shows that a rapid transition to renewable energy sources by 2050 could save the global economy trillions of dollars compared to both a gradual transition and to no transition at all.

Read More

Advancing science with machine learning

At the crossroads of computer science and computational science, the emerging field of scientific machine learning focuses on harnessing new ideas in machine learning together with predictive physics-based models to solve complex, real-world problems. On October 10–12, a group met to collaborate on new ideas about using scientific machine learning in complex fields.

Read More