Inter-Task Interference

This occurs when the latent feature representations learned for Task A are directly modified or overwritten during training on Task B. The model's hidden layers are forced to reconfigure to accommodate new data patterns, erasing the functional mappings necessary for the original task.

• Example: A model trained first on animal classification and then on vehicle classification may lose the ability to distinguish 'cat' from 'dog' as its convolutional filters shift to detect wheels and headlights.
• The interference is most severe when tasks are semantically dissimilar and compete for the same representational resources within a fixed-capacity network.

Interference manifests as a direct conflict in the parameter update directions required for different tasks. The optimal gradient for minimizing loss on a new task points in a direction that increases the loss on a previous task.

• This is formally analyzed using the gradient inner product; negative alignment indicates conflict.
• Methods like Gradient Episodic Memory (GEM) explicitly project new task gradients into a space that does not increase past task losses.
• In multi-task learning, gradient conflict is a primary challenge for finding a single shared parameter update that benefits all tasks simultaneously.

A neural network has a finite parameteric capacity. As tasks are learned sequentially, the network's resources become saturated. Learning new patterns necessitates the re-use of existing weights, leading to catastrophic forgetting of earlier knowledge.

• This is a core challenge of the stability-plasticity dilemma: maintaining old knowledge (stability) while integrating new information (plasticity).
• Parameter-Efficient Fine-Tuning (PEFT) methods, like task-specific adapters, mitigate this by allocating isolated, small parameter subsets to each task, leaving the core model weights frozen and stable.

A critical bottleneck in class-incremental learning where new classes are added over time. The final classification layer must be expanded for new classes, causing direct competition for the logit space.

• Old class logits can be suppressed or 'forgotten' if not rehearsed, as the softmax function normalizes across all classes.
• Common mitigation strategies include using a dynamic output head, bias correction, or knowledge distillation from an old model to preserve old class logit distributions.

Adaptive optimizers like Adam maintain internal state (e.g., momentum, variance estimates) per parameter. When switching tasks, this state is optimized for the previous task's loss landscape and becomes misaligned for the new task, leading to inefficient or destructive updates.

• The optimizer's momentum may carry parameters in a direction beneficial to Task A but harmful to Task B.
• Continual learning strategies sometimes involve resetting or partitioning optimizer states per task to reduce this form of interference.

Interference is exacerbated during inference if the model cannot identify which task an input belongs to. Without a clear task identifier, the model applies a blended set of weights, leading to poor performance on all tasks.

• Task-agnostic learning is the most challenging setting, requiring the model to infer the task.
• Task-aware inference (providing a task ID) allows the model to select the correct task-specific adapter or modulation, isolating parameters and preventing interference at inference time. PEFT excels in this setting by enabling efficient multi-model deployment via lightweight task modules.

Catastrophic forgetting is the abrupt and drastic loss of a neural network's performance on previously learned tasks when it is trained on new data. It is the primary observable symptom caused by inter-task interference.

• Mechanism: Occurs when gradient updates for a new task overwrite the weight configurations critical for old tasks.
• Contrast with Interference: While interference describes the competitive process of learning, catastrophic forgetting is the outcome—the significant performance drop.

The stability-plasticity dilemma is the fundamental challenge in continual learning of balancing the need to retain old knowledge (stability) with the ability to learn new information (plasticity).

• Stability prevents inter-task interference but can lead to intransigence (inability to learn new tasks).
• Plasticity enables learning but risks catastrophic forgetting.
• PEFT Role: Methods like Elastic Weight Consolidation explicitly model this trade-off by selectively constraining parameter updates.

Elastic Weight Consolidation is a regularization-based continual learning algorithm that mitigates inter-task interference by slowing down learning on parameters deemed important for previous tasks.

• Core Idea: Adds a quadratic penalty to the loss function based on the Fisher information matrix, which estimates each parameter's importance to prior tasks.
• Mechanism: Parameters with high Fisher importance are "anchored," making them less plastic and reducing interference.
• PEFT Connection: EWC is a form of selective fine-tuning that imposes a soft mask on the full parameter set, contrasting with hard parameter isolation in adapter-based PEFT.

Gradient Episodic Memory is a constrained optimization method that directly prevents inter-task interference by ensuring gradient updates for a new task do not increase the loss on past tasks.

• Core Idea: Stores a small episodic memory of examples from previous tasks. When computing gradients for a new task, it projects them onto a direction that does not increase the loss on these stored examples.
• Advantage: Provides a formal guarantee against negative backward transfer.
• Limitation: Requires storing raw data, which may not be feasible in all privacy-sensitive or storage-constrained scenarios.

Experience replay is a rehearsal-based technique that mitigates inter-task interference by periodically retraining the model on a stored subset of data (replay buffer) from previous tasks.

• Mechanism: Interleaves examples from the current task with replayed examples from old tasks during training. This simultaneous exposure helps the model find weight configurations that satisfy multiple tasks.
• Variants: Includes generative replay, where a separate model synthesizes pseudo-data for past tasks, eliminating the need to store raw data.
• PEFT Synergy: Often combined with PEFT; a small set of adaptable parameters (e.g., adapters) is tuned on the mixed stream of new and replayed data.

Task-specific adapters are a PEFT architecture strategy to prevent inter-task interference by isolating learned changes into distinct, small neural modules for each task.

• Core Idea: A frozen base model is augmented with lightweight adapter layers. A unique adapter is trained and stored for each task, leaving the shared base model unchanged.
• Interference Mitigation: During inference, only the adapter for the target task is activated, creating a clean functional pathway. This parameter isolation avoids the competition that causes interference in sequential fine-tuning.
• Trade-off: Requires task identity at inference and storage for N adapters, leading to research into adapter composition and task routing.