Optimal decision making
Exploring how movement, beliefs, and spatial geometry shape animal decisions.
In collaboration with Iain Couzin · MPI-AB / University of Konstanz.
I’m a researcher and engineer working at the intersection of natural and artificial intelligence, data science, and animal behavior.
My research explores how natural and artificial systems learn, make decisions, and adapt under uncertainty. Drawing on Bayesian statistics, causal inference, machine learning, and reinforcement learning, I develop theory and models to generate and test hypotheses about the mechanisms underlying behavior.
My previous work spans animal navigation, collective motion, and computer vision tools to measure behavior in the lab and in the wild. I’ve worked with neuroscientists, biologists, physicists, and engineers to develop questions, experiments, and models together. This research has been featured in Quanta Magazine, Nature Methods, Nature, Science, Physics World, National Geographic, and Discover.
I completed my PhD with Iain Couzin at the Max Planck Institute of Animal Behavior and the University of Konstanz, and my MSc with Dan Wiegmann and Vern Bingman at Bowling Green State University.
© M.L. SmithExploring how movement, beliefs, and spatial geometry shape animal decisions.
In collaboration with Iain Couzin · MPI-AB / University of Konstanz.
Developing Bayesian circular statistics to model directional data and quantify uncertainty in animal behavior.
In collaboration with James Foster · University of Konstanz.
Studying how zebras combine terrain, social context, and internal state when choosing where to move.
In collaboration with Blair Costelloe · Mara Elephant Project.
Open model
Photo © Peter Marting
Using a 4D atlas of 12 honeybee nests, we tested how well neighboring cells predict where bees raise brood and store food.
At Auburn, our team photographed both sides of each comb in 12 colonies from April to November 2021. We reconstructed the nests in three dimensions at each observation so we could follow individual cells over time. Cells and even whole combs specialized in raising brood or storing food. In undisturbed nests, pollen accumulated on the side where the nest began and nectar on the opposite side. By fall, stores shifted toward the center. We rearranged the combs in six colonies. Their cells changed use more often and more young brood died, yet worker populations and nest sizes remained comparable to those in undisturbed colonies.
I developed and fitted an autoregressive model that predicts a masked cell’s contents from its neighbors. I compared predictions using neighbors along one, two, or all three spatial axes. The model performed best when it could see across neighboring combs as well as within a comb. Shuffling combs physically or in simulation largely erased that advantage, especially for pollen and nectar or honey. I visualized these results in Figures 6 and S11, reproduced analyses from Peter Marting’s notebooks, and built the interactive explainer.
Artist and biologist Peter Marting turned nest maps from Michael Smith’s lab into Hives Alive, an outdoor installation that opened in September 2026 at Auburn’s Donald E. Davis Arboretum. Up to 2,000 illuminated globes hung from the trees. Each globe represented a cell, and its color showed where the bees were raising young, storing pollen or honey, or leaving cells empty. Photographs and interviews from Auburn University.
In the opening image, the nest photograph dissolves into a 3D view of colony DD1 at its first observation. The colored squares on the parallel comb faces show cell contents recorded in the interactive atlas. Photograph © Peter Marting, cropped as shown in Figure 1A(i) of Marting et al., Current Biology (2026), CC BY 4.0.
By playing, you allow YouTube (Google) to process data and use cookies. Privacy.
Testing how individual decisions give rise to collective motion in locusts.
This study revealed that locust swarming behavior emerges from internal decision-making processes rather than alignment with nearby individuals.
I collaborated with colleagues to analyze behavioral data from controlled lab and immersive virtual reality experiments, developing a Bayesian model to test competing mechanistic hypotheses of collective motion. Our results challenged classical alignment-based theories, showing that group coordination emerges from individual decision-making informed by neural representations of neighboring locusts.
By playing, you allow YouTube (Google) to process data and use cookies. Privacy.
Reconstructing individual trajectories and environmental context with computer vision.
This work introduced a drone- and computer vision–based system for high-resolution tracking of animal groups in the wild.
I helped build and refine a scalable pipeline that combines drone video with deep learning and computer vision algorithms to detect, track, and reconstruct 3D movement and environmental context for wild animals. Working with an interdisciplinary team, we enabled large-scale, non-invasive behavioral analysis in natural settings — advancing tools for behavioral ecology and conservation.
By playing, you allow YouTube (Google) to process data and use cookies. Privacy.
How fish coordinate their tailbeats to exploit the vortices produced by their neighbors.
This paper identified a schooling strategy where fish synchronize their tailbeats to exploit hydrodynamic advantages, matching behavior previously seen in bio-inspired robots.
I contributed to the analysis pipeline, applying deep learning–based pose estimation to track fish movement and test predictions from robotic models. The findings showed that real fish use similar vortex-phase matching strategies to gain energetic benefits during coordinated swimming.
Developing fast, accurate methods for measuring animal posture with deep learning.
DeepPoseKit introduced methods for fast, accurate animal pose estimation, combining multi-scale deep learning with efficient body-landmark localization.
I led the development of DeepPoseKit’s methods, combining multi-scale deep learning, representations of the relationships between body landmarks, and fast landmark localization with subpixel precision. The work addressed the trade-off between speed and accuracy in measuring animal posture across species.
Using lightweight barcode backpacks and computer vision to map individual movements and social networks.
This system uses barcode-equipped backpacks to enable automated, individual-level tracking of birds for behavioral research.
As part of a collaborative project, I helped design and implement the computer vision pipeline used to detect and track barcoded zebra finches. The system provides continuous identity, position, and orientation data, supporting long-term studies of behavior and social networks at high resolution.
By playing, you allow YouTube (Google) to process data and use cookies. Privacy.
Exploring how whip spiders navigate using touch and smell.
Our studies revealed that whip spiders use tactile and olfactory cues — not vision — to navigate complex environments with impressive spatial precision.
I led experimental design and data collection across both lab and field settings, collaborating with colleagues to combine radio telemetry, behavioral tracking, and sensory manipulations. Together, we showed that antenniform legs provide essential multisensory input for goal-directed navigation in the absence of visual cues.