February 2024 Summaries

Scientists at Hume are using semantic space theory (SST) to study emotions in a more comprehensive way than traditional methods. SST uses computational methods and data-driven approaches to map the full spectrum of human feelings, capturing complex, nuanced appraisals that people make about emotion. The approach has led to three major insights: 1) Emotion is high-dimensional, with studies consistently finding that upwards of 20 distinct dimensions characterize emotions; 2) Emotion categories are not discrete but heterogeneous and often blended; and 3) Specific emotions are real and they organize emotional behavior. SST harnesses advanced computational, statistical, and data collection approaches to provide a path forward for the scientific study of emotion that acknowledges the full complexity of human experiences and feelings.

A recent study used machine learning to analyze facial expressions from around the world, finding that there are 28 distinct kinds shared across cultures. The AI model was able to differentiate these expressions and quantify how many had similar meanings across countries. This research provides a detailed understanding of the universality of facial expressions and demonstrates how deep neural networks can be used to investigate psychological processes while controlling for human biases.

HumE-1 (Human Evaluation 1) is a new evaluation method for large language models (LLMs) that focuses on human ratings to assess their ability to perform creative tasks in ways that matter to us, evoking the desired feelings. LLMs are already being used in various fields such as writing books and articles, assisting legal professionals and healthcare practitioners, and providing mental health support. However, existing benchmarks fail to capture how these models affect our satisfaction and well-being. HumE-1 evaluates LLMs on tasks like writing motivational quotes, interesting facts, funny jokes, beautiful haikus, charming limericks, scary horror stories, appetizing descriptions of food, and persuasive arguments for charity donations. The evaluation uses honest and naturalistic prompts to reflect real-life scenarios better. In the first round of results, Gemini Ultra performed best, followed by GPT-4 Turbo, with both models having significant room for improvement.