UNITI: Framework for multi-task learning across datasets to mitigate overfitting

Multi-task learning (MTL) has emerged as a promising approach for improving generalization across related tasks by leveraging shared representations. However, existing MTL techniques primarily focus on intra-dataset learning, where tasks share a common dataset with multiple labels. This approach often fails to address real-world scenarios where tasks originate from heterogeneous datasets, leading to catastrophic forgetting and feature interference. To overcome these limitations, we propose UNITI (Unifying Neural with Inter-dataset for Task Integration), a novel inter-dataset multi-task learning framework that enables a single model to effectively learn from multiple distinct datasets. UNITI consists of two key components: (1) sequential dataset training, which reduces interference by updating model parameters systematically across datasets, and (2) feature-level knowledge distillation (KD), where a student model learns essential task-related features from dataset-specific teacher models. We validate UNITI using both CNN-based (ResNet50) and ViT-based (SHViT) architectures on facial recognition tasks (age estimation, emotion classification) and general object classification (Caltech-101). Experimental results show that UNITI achieves up to 5.17% improvement in accuracy over standard MTL methods and maintains comparable performance to single-task models while significantly reducing computational overhead. Notably, in emotion recognition, UNITI improved accuracy from 60.67% (single-task) to 65.84%, demonstrating its ability to preserve task-specific features in an inter-dataset setting. Our findings suggest that UNITI is a scalable and efficient alternative to traditional MTL approaches, with applications in real-world AI systems where diverse datasets must be integrated into a single model.