EVE: A Task-Adaptive Optimization Algorithm for Balanced Multitask Learning

Optimization plays a central role in multitask learning (MTL), yet widely used optimizers such as Adam and AdamW assume uniform parameter updates across tasks, often leading to optimization bias where simpler objectives dominate training. This paper proposes Extensive Variational Estimation (EVE), a task-adaptive optimizer that extends AdamW by integrating exponential moving average–based task-specific learning rate modulation, scheduled decay of shared parameters, and task-indexed parameter management. Unlike existing optimizers that remain task-agnostic, EVE dynamically adjusts updates according to per-task progress, thereby preserving stability across heterogeneous objectives. The algorithm was empirically validated on three representative multitask domains: image-level classification with pixel-level segmentation on the BUSI breast ultrasound dataset, classification with survival analysis on the SEER prostate cancer cohort, and semantic segmentation with depth estimation and surface normal on the NYUv2 dataset. With a higher Dice coefficient (0.795 ± 0.013) and competitive classification accuracy, EVE performed better on BUSI than Adam and AdamW. On SEER, EVE demonstrated stable performance, improved calibration, and a lower integrated Brier score (0.471 ± 0.035). EVE achieved semantic segmentation performance comparable to Adam (mIoU ≈ 0.370) and higher than AdamW, while maintaining depth estimation performance close to the baselines and exhibiting reduced variance across seeds, on NYUv2. Ablation and Sensitivity experiments proved the importance of EVE's smoothing and minimum learning rate components, and statistical testing revealed consistent directional advantages over Adam in BUSI and SEER, and over AdamW in NYUv2. The proposed optimizer is applicable to multitask learning settings where task heterogeneity and imbalance challenges exist.