Ph.D. (Engg) : Enhancing Precise Label Prediction and Imbalance Robustness in Multi-Label Learning

Multi-label learning (MLL) addresses learning problems in which a single data instance may simultaneously belong to multiple semantic categories. This formulation arises naturally in many real-world applications, including image and video understanding, medical diagnosis, text categorization, and bioinformatics. In many of these settings, it is not sufficient to merely rank relevant labels higher than irrelevant ones; instead, models must accurately identify the exact set of labels associated with each instance. Such exact label prediction is critical when each label carries direct semantic or operational meaning, for example when detecting disease conditions in medical data or identifying pedestrians and traffic signs in autonomous driving scenes. In addition, real-world deployments frequently expose multi-label models to out-of-distribution (OOD) inputs caused by domain shifts, novel concepts, or evolving environments, making reliable OOD detection an important requirement. Many practical applications are also sequential in nature, where data and label spaces evolve over time, leading to the continual multi-label learning (CMLL) problem in which models must acquire new knowledge while mitigating catastrophic forgetting. Together, these considerations motivate the need to study multi-label learning with emphasis on exact label prediction, robustness to data imbalance, improved OOD detection, and learning under continual data arrival.

The first contribution of this thesis introduces Bipolar Networks, a novel architectural formulation for multi-label classification designed to improve exact label prediction. Unlike conventional single-output architectures that produce continuous confidence scores per label, Bipolar Networks represent each label using two complementary outputs that encode positive and negative evidence. The final label decision is derived from the relative difference between these outputs, enabling exact label predictions. To support effective training of this architecture, the thesis develops a family of bipolar loss functions by reformulating standard objectives such as Binary Cross-Entropy and Focal Loss, along with margin-based variants. Extensive experiments on benchmark datasets demonstrate that Bipolar Networks consistently improve F1 scores while maintaining competitive mean average precisions.

Building on the improved discriminative behaviour of Bipolar Networks, the second contribution addresses out-of-distribution detection in multi-label learning. While most existing OOD detection methods are designed for single-label classification or rely on computationally intensive mechanisms, this thesis proposes a bipolar joint energy score tailored to the bipolar architecture. By leveraging the improved exact label prediction capability of Bipolar Networks, the proposed scoring function enables more effective separation between in-distribution and out-of-distribution samples in multi-label settings, demonstrating that stronger multi-label classification performance on in-distribution data can naturally translate into improved OOD detection.

The third contribution presents Learn What Matters, a generalizable training framework that enhances exact label prediction without modifying model architectures or loss formulations. Learn What Matters operates at the optimization level by selectively masking parameter updates based on the ratio of gradient to parameter magnitudes, suppressing low-information updates while rescaling the remaining gradients to preserve learning dynamics. This approach acts as a form of dropout during backpropagation and directs learning toward informative regions of the parameter space. Applied to standard single-output multi-label networks trained with ranking-based losses, Learn What Matters yields substantial improvements in F1 score with only marginal impact on mAP scores, providing a model-agnostic alternative to architectural modifications.

The fourth contribution explores biologically inspired learning approaches for multi-label classification. Drawing inspiration from neural computation in the human brain, the thesis develops several bio-inspired models, including Bipolar Spiking Neural Networks, Adaptive Margin Spiking Neural Networks, and NIMBLE, a neuro-inspired multi-label learning framework. These methods leverage spike-based computation, selective update mechanisms, and adaptive stability–plasticity behavior to naturally support exact label prediction and efficient learning. Experimental results demonstrate that these biologically motivated designs improve exact label prediction and imbalance robustness in multi-label settings.

Finally, the thesis extends the proposed architectures and learning algorithms to the continual multi-label learning setting, where data arrives sequentially without access to past samples or task identities. The resulting methods are task-agnostic and memory-free, and empirical evaluations across multiple benchmarks show consistent improvements in exact label prediction and reduced catastrophic forgetting. Overall, this thesis presents a unified framework spanning architectural design, optimization strategies, and biologically inspired learning for advancing exact label prediction, OOD detection, and continual learning in multi-label learning models.

Speaker : Sourav Mishra

Research Supervisor: Prof. Suresh Sundaram