Adaptive and Responsible Machine Learning: From Incremental Learning to Unlearning and Safe Generation

Modern machine learning (ML) systems are increasingly required to operate in dynamic environments, where models must adapt to new information, remove obsolete or sensitive knowledge, and satisfy fairness and safety requirements. This thesis addresses these challenges by developing mechanisms for controlling what a model learns, retains, and forgets. The first part studies incremental learning (IL) in data-free and class-incremental settings. To mitigate catastrophic forgetting and inter-task confusion, an embedding-based regularization strategy is proposed that separates current and previously learned class representations using stored class centroids. In addition, a generative replay mechanism reusing a previously trained model synthesizes representative samples of past tasks. A knowledge distillation formulation operating on embedding clusters is further developed to preserve intra-class and inter-class structures, enabling scalable IL without storing historical data. The second part investigates machine unlearning in discriminative models under strict data-free constraints. A new unlearning framework, termed MaxDiv, is proposed to remove the influence of designated training data by generating erasing samples that reverse the feature distribution of the data to be forgotten. The proposed method combines erasure-oriented objectives with representation-preserving losses to achieve effective forgetting while maintaining performance on remaining data. Experiments demonstrate improved forgetting guarantees, resistance to model inversion attacks, and superior preservation of retained knowledge compared to existing unlearning approaches. The third part extends machine unlearning to generative models, where the goal is often to control undesirable behaviors such as bias, unsafe content, and ethical misalignment. Following the responsible AI paradigm, this thesis studies responsible image generation in text-to-image (T2I) diffusion models (DMs) as a representation-level intervention problem. We first focus on U-Net-based DMs, where precise concept vectors are learned in the bottleneck layer to generate responsible images that preserve foreground and background content. We then demonstrate that multi-modal attention heads in ViT-based DMs exhibit interpretable and near-linear properties analogous to the bottleneck layer in U-Net-based architectures. Building on this observation, we introduce an architecture-agnostic framework for learning interpretable concept vectors that enable controllable, disentangled, and human-interpretable manipulation of internal representations across U-Net- and ViT-based diffusion backbones. Overall, this thesis provides a unified framework for adaptive and responsible ML.