Intelligent Predictive Frameworks Under Data Scarcity And Uncertainty

Modern predictive systems frequently operate under conditions of limited annotated data, high uncertainty, and the need for reliable decision-making. When the predictive models expand across heterogeneous data types (e.g., spatial, temporal streams), the challenge lies not only in accurate prediction but also in adapting in data distributions shifts or label scarcity. To address these issues, this thesis explores an Intelligent Predictive Framework that operates robustly under data scarcity and uncertainty across two distinct domains: medical imaging (spatial) and time-series forecasting (temporal). In the first part of this work, a semi-supervised mean teacher (MT) paradigm is tailored for medical image segmentation under limited supervision. Furthermore, to generalize the paradigm beyond imaging data, we extended its application in time series forecasting. In the medical imaging setting, the proposed framework achieved notable segmentation accuracy compared to supervised baselines. Even under limited annotation (e.g., using only 80% of the labeled data), it reached a dice coefficient of 0.613 ± 0.01 and a jaccard index of 0.441 ± 0.02, outperforming the supervised counterparts. In the same way, the adapted model produced measurable gains in time series domain, achieving a 25.4% reduction in mean squared error (MSE) compared to the baseline student model. Despite significant progress, the current framework remains limited in their ability to capture long-range temporal structures in cross-domain platforms. Revealing those clear gaps, the thesis further outlines a proposal for a biologically inspired time series foundation model (BioPhant-TSFM). Motivated by the brain’s cortical–hippocampal memory system and elephants’ exceptional long-term episodic recall, the proposed BioPhant-TSFM integrates prototype memory, and episodic retrieval to support long-range forecasting and extreme-event reasoning. We further aim to deploy the proposed model on a Unitree robot dog platform conceptualized as Memophant, enabling real-world long-horizon forecasting and adaptive memory consolidation in a mobile embodied system.