Identification of Abnormal States in Videos of Ants Undergoing Social Phase Change

Taeyeong Choi, Benjamin Pyenson, Juergen Liebig, Theodore P. PavlicAAAI 2021 · DOI

Abstract

Quantifying transitions from order to disorder in complex social systems is a significant challenge. This study presents a novel one-class classification framework, IO-GEN, specifically designed to detect "abnormal" social behaviors in colonies of Harpegnathos saltator (Indian jumping ants). By training exclusively on "normal" stable data and utilizing a unique "inner-outlier" generation strategy, the model successfully identifies the subtle, chaotic transitions that occur when a colony loses its queen and must reform its social hierarchy.

Introduction: The Challenge of Social Phase Changes

In social insect colonies, transitions between stable and unstable states are often difficult to pinpoint. For Harpegnathos saltator, the death or removal of a queen triggers a ~20-day "unstable" phase where workers compete to become "gamergates" (reproductive individuals). This period is characterized by aggressive dueling and dominance biting. While these behaviors are observable to human experts, automating their detection is a classic One-class Classification (OC) problem: stable data is abundant, but unstable data is rare, diverse, and difficult to label precisely.

Methodology: Inner-Outlier Generation (IO-GEN)

The proposed framework builds upon Deep Support Vector Data Description (DSVDD), which maps data into a feature space where normal samples cluster around a central point. The core innovation of this work is the Inner-Outlier Generator (IO-GEN).

Unlike standard generative models that attempt to mimic the entire distribution of normal data, IO-GEN is constrained to generate "inner outliers"—synthetic samples that represent the most prototypical, ultra-stable behavioral patterns. By training a binary classifier to distinguish between real stable data (labeled as "unstable" for this step) and these synthetic "super-normal" samples, the model is forced to learn the highly specific features that define the most stable states. This increased sensitivity allows the model to detect even subtle deviations during the colony's transition phase.

Visual comparison of feature space distributions. (a) Standard Autoencoder, (b) DSVDD, (c) DSVDD with synthetic outliers, and (d) the proposed IO-GEN approach which focuses on the center of the hypersphere.

To ensure the model learns behavioral dynamics rather than static scene features, the researchers utilized optical flow sequences instead of raw RGB pixels. This approach focuses the model on the patterns of motion between frames, which serve as a proxy for social interactions.

Experimental Results and Discussion

The model was evaluated on a custom dataset of an H. saltator colony recorded over 20 days, with a social phase change triggered by the removal of gamergates between Day 2 and Day 3.

Performance and Reliability

IO-GEN achieved an AUC of 0.786, outperforming standard DSVDD (0.762) and traditional One-Class SVM (0.523). The model proved particularly effective during the middle of the transition period (Days 6–10), where the colony begins to stabilize but still exhibits cryptic aggressive interactions.

Anomaly scores tracked over the 20-day experimental period. The red vertical line marks the removal of the queen. Scores spike immediately following the event and gradually decrease as the colony returns to a stable state.

As shown in the temporal analysis of anomaly scores, the model maintains a clear detection margin throughout the transition. While the scores naturally decrease as the colony re-establishes its hierarchy, the IO-GEN framework remains more sensitive to lingering instabilities compared to baseline methods.

Conclusion and Future Work

The IO-GEN framework demonstrates that defining "perfectly normal" behavior through synthetic inner outliers is a powerful strategy for anomaly detection in complex social systems. By leveraging optical flow, the model captures essential social dynamics without the need for individual ant tracking.

Future work will explore the generalizability of this approach to other social species and different types of transitions, such as panic in human crowds. Additionally, the authors aim to improve interpretability by developing visualization tools that highlight the specific regions of a video frame contributing to an "abnormal" classification, aiding biologists in identifying specific behavioral triggers.

Summary updated

BibTeX
@inproceedings{choi2021identification,
  author = {Taeyeong Choi and Benjamin Pyenson and Juergen Liebig and Theodore P. Pavlic},
  title = {{Identification of Abnormal States in Videos of Ants Undergoing Social Phase Change}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17794},
}
Figure