Contrastive Learning for Comparative Behavioural Analysis - Extracting Behavioural Features from Rat Trajectory Data

Abstract

Abstract When conducting drug development research, movement pattern analysis is a valuable approach for examining behavioural variations in rats subjected to different substances. In order to reduce the risks of human bias and missed details associated with manually engineered models, machine learning is a viable option to find behavioural features directly from the data. Contrastive learning models constitute one such method, which can learn to find relevant behavioural aspects by representing similar substance-induced behaviours similarly. In this thesis, we develop a deep neural network which utilizes contrastive learning to extract behavioral features from rat trajectories induced by different substances. Additionally, various model variations are evaluated and compared against an existing model based on manually engineered features. The results demonstrate similar performance between the proposed and manually engineered models. Surprisingly, the proposed model displays insensitivity to different modifications, and the application of techniques proven successful by other contrastive learning studies does not further enhance performance. These findings suggest a potential underlying issue that may stem from the data, learning approach, or chosen model architecture.