Class-Incremental Learning

Catastrophic forgetting is the primary technical obstacle, where a neural network's performance on previously learned classes degrades dramatically when trained on new data. This occurs because gradient-based optimization for new tasks overwrites the weights critical for old tasks. In class-incremental learning, this manifests as a sharp drop in accuracy for old classes as the model adapts to new ones. Mitigation is the central goal of continual learning algorithms.

This is the fundamental trade-off at the heart of continual learning. Stability refers to a model's ability to retain knowledge of old classes. Plasticity is its capacity to learn new classes effectively. Naive fine-tuning maximizes plasticity at the expense of stability, leading to forgetting. Excessive regularization for stability can stifle learning of new concepts. Class-incremental algorithms must dynamically balance this tension across the entire learning sequence.

The final classification layer presents a major architectural challenge. A fixed-size output layer cannot accommodate new classes. Common solutions include:

• Expanding the classifier by adding new output neurons for each new set of classes.
• This can create a bias toward newly learned classes, as their associated weights are freshly optimized while old weights may have stagnated.
• Techniques like weight aligning or bias correction during inference are often required to re-balance logits across all classes, old and new.

A core practical challenge is the prohibition of storing or revisiting old training data due to privacy, storage, or regulatory constraints. When a small replay buffer or exemplar memory is allowed, key issues are:

• Exemplar selection: Choosing which old examples to retain for maximum retention benefit.
• Memory management: Efficiently using a fixed memory budget across a growing number of classes.
• Rehearsal strategies: Determining how and when to interleave old exemplars with new task data during training.

Unlike task-incremental learning, class-incremental learning is task-agnostic at inference time. The model is not provided with a task identifier and must discriminate among all classes seen so far from a single, unified output space. This eliminates the simple strategy of using a task-specific classifier head and forces the model to develop a shared, disentangled feature representation that can separate all classes without explicit task context, a significantly harder problem.

Parameter-Efficient Fine-Tuning methods like LoRA or Adapters offer a natural architectural advantage for class-incremental learning. By freezing the base model and learning small, task-specific modules, they inherently reduce inter-task interference. For each new set of classes, a new adapter can be trained while old adapters remain frozen. The challenge shifts to designing adapter composition or routing mechanisms at inference time to activate the correct knowledge without a task ID.

PEFT directly combats catastrophic forgetting—the tendency of neural networks to lose performance on old tasks when learning new ones. By freezing the vast majority of the pre-trained model's parameters and updating only a small, task-specific subset (e.g., LoRA matrices or adapters), the foundational knowledge encoded in the base model remains largely intact. This provides inherent stability. The small, learnable parameters provide the necessary plasticity to acquire new class discriminators without overwriting the general representations needed for old classes.

In class-incremental learning, the model must consolidate knowledge of an expanding set of classes. PEFT facilitates this by learning compact, modular representations for each new set of classes. Techniques like task-specific adapters or prefix tuning allow the model to develop isolated parameter spaces for new knowledge. During inference, the correct module is activated, allowing the model to leverage both the frozen, general-purpose base model and the specialized, efficient module for the current task, enabling efficient recall without interference.

A naive approach of sequential fine-tuning the entire model for each new set of classes is computationally prohibitive and leads to forgetting. PEFT offers a scalable alternative. For each new batch of classes, only the small PEFT parameters are trained, while the base model serves as a stable, shared backbone. This results in:

• Linear parameter growth: Storage scales with the number of tasks/class groups, not the model size.
• Independent deployment: PEFT modules can be saved, swapped, or composed.
• Reduced compute cost: Training is drastically faster and cheaper than full fine-tuning.

PEFT is highly compatible with experience replay, a common class-incremental learning technique. When replaying old data from a replay buffer, only the small PEFT parameters associated with those old classes need to be updated, making the rehearsal process extremely efficient. Furthermore, PEFT can be combined with generative replay, where a generative model creates synthetic examples of past classes. The main classifier, stabilized by its frozen backbone, uses its efficient PEFT parameters to learn from this synthetic data, preventing drift.

The stability-plasticity dilemma is central to continual learning: how to stay plastic enough to learn new things but stable enough to remember old ones. PEFT architectures provide an engineered solution:

• Stability: The frozen, massive pre-trained model provides a stable, general-purpose feature extractor.
• Plasticity: The small, trainable PEFT parameters (often <1% of total parameters) offer a dedicated, malleable subspace for new learning. This separation allows the model to be highly plastic within a confined, efficient parameter space while the core model remains stable, offering a more controlled balance than tuning the entire network.

For production systems, PEFT makes class-incremental learning operationally feasible. Key advantages include:

• Lightweight Updates: Deploying a new model version often means distributing only a few megabytes of adapter weights instead of a multi-gigabyte full model.
• Rapid Iteration: New classes can be added with quick training cycles on limited data.
• Multi-Tenancy: A single base model can host multiple, isolated PEFT modules for different clients or use cases, all running on the same inference infrastructure.
• Backward Compatibility: The core model remains unchanged, simplifying versioning and rollback procedures.

Continual learning is the overarching machine learning paradigm where a model learns from a non-stationary stream of data, acquiring new knowledge from sequential tasks while aiming to retain performance on previous ones. It is the parent field that encompasses class-incremental learning.

• Core Objective: Accumulate knowledge over a lifetime without catastrophic forgetting.
• Key Scenarios: Includes task-incremental, domain-incremental, and class-incremental learning.
• Primary Challenge: Solving the stability-plasticity dilemma—balancing the retention of old knowledge with the acquisition of new information.

Catastrophic forgetting is the severe and abrupt degradation of a neural network's performance on previously learned tasks when it is trained on new data. It is the primary obstacle that continual learning methods, including those for class-incremental learning, are designed to overcome.

• Mechanism: Occurs due to inter-task interference during gradient-based optimization, where weight updates for a new task overwrite representations crucial for old tasks.
• Analogy: Like a student mastering calculus but then completely forgetting basic algebra after a new lesson.
• Mitigation: Addressed by techniques like experience replay, regularization (e.g., EWC), and parameter-efficient fine-tuning (PEFT) to isolate or protect important parameters.

Experience replay is a foundational continual learning technique where a model rehearses on a stored subset of data from past tasks while learning a new task. It directly combats catastrophic forgetting by re-exposing the model to old patterns.

• Core Component: Uses a replay buffer (or memory) to store a limited number of exemplars from previous classes or tasks.
• Variants: Includes generative replay, where a separately trained generative model produces synthetic data of past tasks for rehearsal.
• Trade-off: Effectiveness is balanced against memory overhead and potential privacy concerns from storing raw data.

Elastic Weight Consolidation is a regularization-based continual learning algorithm that estimates the importance of each model parameter to previously learned tasks and penalizes changes to important parameters during new training.

• Mechanism: Calculates a Fisher information matrix to approximate parameter importance. It then adds a quadratic penalty term to the loss function to constrain updates to critical weights.
• Analogy: Treats important neural connections as "springs" that resist change.
• Relation to PEFT: EWC is a soft, regularization-based approach to parameter efficiency, whereas PEFT methods like adapters enforce a hard, architectural sparsity by freezing most parameters.

Task-agnostic learning is a challenging continual learning setting where the model receives no explicit signal about task boundaries or identities during training or inference. The model must autonomously detect distribution shifts and manage its own knowledge.

• Contrast with Class-Incremental: In class-incremental learning, task boundaries (when new classes arrive) are typically known during training, but the task ID is not provided at test time. Task-agnostic is a stricter, more realistic but difficult scenario.
• Challenge: Requires sophisticated mechanisms for online learning, change detection, and internal knowledge organization without external guidance.

Parameter-efficient fine-tuning is a family of techniques that adapts large pre-trained models by updating only a small, strategically selected subset of parameters (e.g., adapters, prefixes, or low-rank matrices), leaving the vast majority of the base model frozen.

• Core Value for Continual Learning: PEFT provides a natural architectural bias against catastrophic forgetting. By isolating task-specific knowledge into small, modular components (like task-specific adapters), interference with the core model is minimized.
• Methods: Includes Low-Rank Adaptation (LoRA), prefix tuning, and adapter modules.
• Application: Enables efficient sequential fine-tuning across multiple tasks or an expanding set of classes, making it highly relevant for class-incremental learning systems.