Decoding Collective Behavior: How Active Bayesian Inference Powers the Natural Movements of Animal Groups

The phenomenon of collected motion in animals observed in activities like swarming locusts, schooling fish, flocking birds, and herding ungulates is extensively studied due to its visually striking property and its emergence from simple interactions among group members. Recent research focuses on more biologically motivated, agent-based approaches that aim to model specific behavioral circuits and decision rules that govern individual behaviors. Researchers have designed a model based on active inferencing that bridges theoretical and biological aspects of human behavior.

This model class unifies cognitive and physics-based perspectives, offering a comprehensive understanding of adaptive behavior. It focuses on how an individual estimates distance to neighbors and uses the details to make decisions. It has two parts – the dynamic model, which describes how distances change over time, and the observation model, which explains how individuals sense these distances. Active inference updates its beliefs and actions to minimize the surprises. 

The model emphasizes how intricate behavior arises from simple actions driven by predictions and sensory output features. In some scenarios, it converges to traditional force vectors like attraction, repulsion, and alignments, derived as free energy functional, acting as the upper limit on the surprise. Behavioral plasticity is a key mechanism that helps to enhance and collectively represent temporary fluctuations. Unlike using additional rules or mechanisms for specific outcomes, plasticity involves conducting gradient descent on free energy for model parameters. This mechanism is integrated into active inference, extending its application to model parameter updates.

The researchers hope that their work will serve as a link between existing theoretical models of collective animal behavior and more neuro/ML-adjacent fields like active inference and the Bayesian brain framework. They also emphasize that their chosen model explains key attributes observed in collective systems and effectively reproduces the capability to enhance and decode the information that earlier models struggled to model without invoking additional mechanisms. 

Check out the Paper and GitHub. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 27k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.