Next Generation Psychological Embeddings

People have vast knowledge bases that allow them to represent relevant information about the world and act upon it. While there are specialist computer systems that best people in specific tasks, such as playing chess, humans are still the champions at being generalists. How do humans represent their rich knowledge so that they can appreciate similarities between objects, whether those similarities rest on superficial properties or deep connections, such as belonging to a shared biological category? It is a difficult question to answer that has both theoretical and practical ramifications. Understanding how people perceive the world is key to predicting and improving human behaviour. Likewise, building such representational or embedding spaces would provide a powerful tool for making AI systems more human-like.

Standard techniques for inferring psychological representations have been in wide use since the 1950s, but are limited in important ways. Standard techniques are data hungry and computationally slow. As a consequence, these techniques do not work well with real-world problems that often contain more than a million items. We aim to help Psychology transition to large-scale modelling, which we hope leads to a revolution like that experienced a decade ago in machine learning and AI when those fields moved to large-scale datasets. Another limitation of standard techniques is that they can't detect or take advantage of the relationships between items in the representational space. For example, if people know that two breeds of dogs are both dogs, even if they differ in size, they use that structural knowledge when making inferences. Our modelling approaches can discover and use such conceptual relationships. Likewise, we can capture how different groups, who vary in their life experiences, may represent the world in slightly different ways. In doing so, we can capture the uniqueness of each human experience rather than force a one-size-fits-all approach on the data, which would be a kind of tyranny of the majority for data science and Psychology. Finally, our methods can be adapted to take advantage of different approaches to measuring similarity.

Collectively, these limitations in standard approaches block the transfer of laboratory insights into real-world settings. While work has been done to address some of these limitations, no work has addressed all these limitations fully at once. We aim to do so at scale considering two databases of natural images (i.e., photographs) that each contain over a million images. Rather than offer an incremental advance, we aim to advance the state-of-the-art for representing spaces by more than an order of magnitude in size and improve the quality of the solution by capturing relations between images and groups of people as discussed above. We will make these resources and the tools publicly and freely available with guidance on how they can be extended to support others' work, whether it be in Psychology, Education, Human-Computer Interaction, AI, or other fields. Inferring psychological representations for these two large datasets will remove a long-standing hurdle in the research community, which should help machine learning and cognitive science researchers create better models of human cognition. This new framework and resource will make it possible to model differences between individuals, allowing us to better understand how different life experiences, such as measured by age, gender, and geographical location, impact how we think about the world.