VIBE: Enhancing Unsupervised Continual Learning with Autonomous Novelty Detection

Humans naturally recognize and integrate novel stimuli, continuously adapting through synaptic plasticity and neuronal mechanisms. In contrast, Deep Neural Networks (DNNs) struggle with unexpected inputs due to their reliance on independent and identically distributed (i.i.d.) data, limiting adaptability in dynamic environments. Traditional continual learning methods often require supervised weight adjustments and cloud-based oracles, making them impractical for disconnected scenarios like remote wildlife monitoring. The development of fully autonomous learning agents capable of integrating continual learning with unsupervised novelty detection—without relying on a novelty oracle—remains largely underexplored. To address this, we propose VIBE, a neuro-inspired enhancement for unsupervised continual learning models to perform autonomous novelty detection. VIBE uses a differential threshold function, to analyze firing patterns and dynamically reallocate resources. This mechanism allows the system to efficiently distinguish between familiar and novel inputs, autonomously reorganizing internal clusters without human supervision or cloud resources. Integrated with Spiking Neural Networks (SNNs), VIBE leverages their temporal dynamics to enhance unsupervised learning in a biologically plausible way. Validated on real-world scenarios with shifting distributions and novel, unlabeled inputs, VIBE achieves significant accuracy improvements—up to 44.6% (29.85% on average)—over traditional unsupervised continual learning models.