Unsupervised Continual Learning From Synthetic Data Generated with Agent-Based Modeling and Simulation: A preliminary experimentation

Continual Learning enables to learn a variable number of tasks sequentially without forgetting knowledge obtained from the past. Catastrophic forgetting usually occurs in neural networks for their inability to learn different tasks in sequence since the performance on the previous tasks drops down in a significant way. One way to solve this problem is providing a subset of the previous examples to the model while learning a new task. In this paper we evaluate the continual learning performance of an unsupervised model for anomaly detection by generating synthetic data using an Agent-based modeling and simulation technique. We simulated the movement of different types of individuals in a building and evaluate their trajectories depending on their role. We collected training and test sets based on their trajectories. We included, in the test set, negative examples that contain wrong trajectories. We applied a replay-based continual learning to teach the model how to distinguish anomaly trajectories depending on the users’ roles. The results show that using ABMS synthetic data it is enough a small percentage of synthetic data replay to mitigate the Catastrophic Forgetting and to achieve a satisfactory accuracy on the final binary classification (anomalous / non-anomalous).