Class-Incremental Learning

Catastrophic forgetting is the primary obstacle, where a neural network's performance on previously learned classes degrades sharply when its parameters are updated to accommodate new ones. This occurs because gradient-based optimization overwrites the weights that encoded knowledge of old tasks. Unlike task-incremental learning, the model cannot rely on a task ID to select specialized parameters, forcing all updates through a shared network. The stability-plasticity dilemma is central here: the model must be plastic enough to learn new classes but stable enough to retain old ones.

A defining constraint of class-incremental learning is that the model is not provided with a task identifier during evaluation. It must correctly classify an input among all classes seen so far without knowing which 'task' or data phase it belongs to. This eliminates simpler solutions like using a multi-head output layer where each head corresponds to a known task. The model must develop a unified, discriminative representation space that separates all classes—old and new—simultaneously, which is significantly harder than task-aware scenarios.

Models suffer from a representation bias toward newly learned classes and a classification bias toward the most recent task. As the model trains extensively on new data, its feature extractor tunes its representations to discriminate within the new class set, often at the expense of separability for old classes. Consequently, at test time, there is a strong tendency to misclassify samples as belonging to a new class. Techniques like knowledge distillation and nearest-mean-of-exemplars classification are used to combat this, but maintaining a balanced decision boundary across an expanding number of classes remains a major challenge.

Since storing all past data is typically infeasible, a small replay buffer or exemplar set is maintained. This introduces critical sub-problems:

• Selection Strategy: Which old samples to retain? Common approaches include herding (selecting prototypes closest to the class mean) or random selection.
• Budget Allocation: How many exemplars per old class? A fixed total budget must be distributed fairly as the number of classes grows, often leading to fewer examples per class over time.
• Rehearsal Scheduling: When and how often to interleave replay of old data with new training? Poor management directly amplifies forgetting and bias.

The model's output layer must grow to accommodate new classes. This requires:

• Dynamic Architecture: Adding new output neurons (heads) for each new class or group of classes.
• Initialization: Sensibly initializing new weights to integrate with the existing classifier without disrupting it.
• Consolidation: Algorithms like iCaRL replace the standard softmax classifier with a nearest-mean rule using exemplars to avoid the need to retrain the entire output layer, which is prone to bias. Managing this expansion efficiently, without a performance drop, is a key algorithmic design consideration.

Rigorous evaluation is complex. Key metrics include:

• Average Incremental Accuracy: The average test accuracy over all classes seen after each learning phase.
• Forgetting Measure: The drop in accuracy on a task from its peak performance after subsequent training.
• Intransigence: The inability to learn new tasks, measuring plasticity loss. Benchmarks like Split CIFAR-100 (10 tasks of 10 classes) or Split ImageNet simulate the incremental setting. Results must be reported over multiple random task orders, and the exemplar memory budget (e.g., 2000 total samples) must be clearly stated for fair comparison.

Catastrophic forgetting is the tendency of a neural network to abruptly and drastically lose performance on previously learned tasks when trained on new data. It is the primary problem that continual learning algorithms, including class-incremental learning, are designed to solve. The phenomenon occurs due to unconstrained parameter overwriting during gradient-based optimization on new tasks, which erodes the representations needed for old tasks.

Task-incremental learning is a continual learning scenario where tasks are presented sequentially, and the model is provided with an explicit task identifier during both training and inference. This is a less challenging setting than class-incremental learning because the task ID acts as a selector, allowing the model to use dedicated sub-networks or heads. The core challenge shifts from pure discrimination to managing parameter isolation and forward transfer between tasks.

Domain-incremental learning is a scenario where the input data distribution (domain) changes over time, but the underlying output classes and task remain the same. For example, a model must recognize digits (0-9) when the data stream shifts from MNIST (handwritten) to SVHN (street view house numbers). The challenge is to maintain a stable decision boundary while adapting to new visual styles, testing the model's feature robustness and representation stability.

Experience Replay (ER) is a foundational technique for mitigating catastrophic forgetting by storing a subset of past training examples in a fixed-size replay buffer. During training on new data, the model interleaves batches from the buffer, rehearsing old patterns. Key engineering decisions include:

• Buffer sampling strategy (e.g., reservoir sampling for online streams)
• Buffer management policy for which old examples to retain
• The rehearsal loss function, often the standard cross-entropy

This technique uses knowledge distillation to preserve a model's existing capabilities. When learning a new task, the model's predictions on new data are guided by two objectives: the standard loss for the new task and a distillation loss that penalizes deviation from the old model's outputs. This creates a soft regularization that encourages the new parameters to remain consistent with the old function's behavior on the input space, even without old data.

The stability-plasticity dilemma is the fundamental trade-off in continual learning systems. Stability refers to the system's ability to retain knowledge of old tasks. Plasticity is its capacity to learn new information efficiently. All continual learning algorithms, including class-incremental methods, navigate this tension:

• Regularization-based methods (e.g., EWC) favor stability by constraining important parameters.
• Replay-based methods (e.g., ER) attempt to balance both by rehearsing old data.
• Architectural methods (e.g., Progressive Nets) maximize plasticity by adding capacity but can reduce parameter efficiency.