A new technique transforms any computer vision model into one that can explain its predictions using a set of concepts a human could understand. The method generates more appropriate concepts that boost the accuracy of the model

By tracking nearly every movement of a tiny fish’s life from adolescence to death, a new study reveals a hidden behavioral blueprint of aging – one that can predict a fish’s age or how long an individual will live. This is possible based on behavioral patterns visible early in life, researchers report. Aging in vertebrates unfolds over long and complex timescales and is influenced by a myriad of factors. Behavior provides a powerful window into an animal’s internal state and has been shown to reflect the aging process in several species, including humans. However, the ability to continuously observe behavior across an organism’s full lifespan has posed a significant challenge to researchers. As a result, the behavioral structure of aging and how early-life behavioral traits relate to lifespan have remained poorly understood.

To overcome this challenge, Claire Bedbrook and colleagues developed a high-resolution, continuous behavioral recording platform to monitor naturally short-lived African turquoise killifish, which have a lifespan of only a few months. Using machine learning and computer vision, the platform tracked killifish behavior from adolescence (~3 to 4 weeks of age) until death to map how behavior changes across adulthood, determine whether behavioral patterns can predict aging and remaining lifespan, and identify distinct stages of adult life. Bedbrook et al. found that individual animals follow distinct aging trajectories, with long-lived and short-lived individuals showing distinct behavioral differences early in life. Specifically, fish that ultimately live longer were more active, faster-moving, and displayed more vigorous bursts of movement than those that die early. What’s more, long-lived individuals confine most of their sleep to the night. Short-lived fish, on the other hand, exhibit increased daytime sleep and more disrupted activity patterns. By applying a machine learning model to these behavioral measurements – collectively called a “behaviorome” – Bedbrook et al. developed a “behavioral clock” that could estimate an animal’s age using only its daily patterns of movement and activity. The model was also able to show that, beginning in early adulthood, behavioral patterns alone could reliably forecast whether a fish would ultimately have a relatively short or long lifespan.

Feeding children ultra-processed foods, such as chicken nuggets, is common in the US. Social norms like this are difficult to change, but Rutgers Health researchers found that when parents take photos of the food available to them, their perceptions shift and they begin to question this norm.