Explaining Loss of Plasticity in Continual Learning via Loss Landscapes and Mitigating it through Expressivity Preservation

In deep learning research, enabling models to continuously and robustly acquire new knowledge represents a crucial step toward achieving more adaptive and efficient systems. Yet, when exposed to sequential streams of data, neural networks often struggle to maintain performance, suffering either from catastrophic forgetting (the erosion of previously acquired knowledge) or from loss of plasticity (LoP). This thesis specifically addresses the loss of plasticity phenomenon, marked by the progressive decline of a neural network’s representational capacity and the resulting inability to learn effectively from new data, particularly when training relies on traditional optimization methods, such as stochastic gradient descent (SGD) combined with backpropagation. Recognizing that evolving data corresponds to changing tasks, and thus to shifts in the loss function and in its landscape, we frame our analysis through a loss landscape perspective and provide a novel characterization of LoP by examining the geometric characteristics of the landscape traversed by the optimizer. Building on these observations and grounded in existing theoretical notions of convergence, we provide a deeper understanding and justification for why the convergence of SGD with backpropagation breaks down as training enters the LoP regime. Furthermore, we re-propose the use in continual learning of an optimizer, which we name ReRankSGD, designed to mitigate LoP through a conceptually simple yet remarkably effective idea. Our approach, which to the best of our knowledge has so far been proposed only in Deep Reinforcement Learning scenarios, proves highly advantageous in counteracting loss of plasticity by preserving the network inner layers’ expressivity, as suggested by empirical evidence.