Catastrophic Forgetting

Unlike human learning, where new information is integrated with old, neural networks experience a sharp, non-gradual degradation in performance on prior tasks. The model's internal representations are overwritten by new training data, causing previously learned patterns to be erased. This is not a slow forgetting but a rapid collapse of performance metrics on the original task, often occurring within the first few epochs of training on the new dataset.

The core mechanism is destructive interference in the model's parameter space. When gradient descent updates the network's weights to minimize loss on a new task (Task B), those updates are not constrained to protect knowledge of the old task (Task A). The optimization path for Task B directly conflicts with the weight configuration optimal for Task A. This is especially severe in highly over-parameterized models where weights are densely entangled across different functions.

The severity of forgetting is heavily influenced by the order and similarity of learned tasks.

• Task Order: Learning Task A then Task B often causes more forgetting of A than the reverse order.
• Task Similarity: Forgetting is less severe when new tasks are semantically or structurally similar to old tasks, as weight updates may be more aligned. Drastically different tasks (e.g., image classification then language translation) cause maximal interference. This makes catastrophic forgetting a critical problem for continual learning systems that must adapt over time.

Ironically, the large, modern neural networks most capable of learning complex tasks are also most susceptible to catastrophic forgetting. Their high dimensionality and redundant parameterizations mean there are many possible weight configurations that solve a given task. Gradient descent can easily find a new solution for Task B that is distant in weight space from the solution for Task A, with no mechanism to preserve the old pathway. This creates a stability-plasticity dilemma: the network must be plastic to learn new things but stable to retain old knowledge.

A proposed solution highlights the characteristic. Elastic Weight Consolidation (EWC) is an algorithm designed to mitigate forgetting by identifying which parameters are most important for previous tasks and making them less plastic (more "elastic") during new training. It calculates a Fisher Information Matrix to estimate parameter importance. This directly counteracts the unconstrained interference by applying a regularization penalty that anchors important weights to their old values, demonstrating that the core issue is a lack of such constraints in standard training.

Parameter-efficient fine-tuning methods like LoRA inherently reduce catastrophic forgetting risk. By freezing the base model and only training small adapter weights (e.g., low-rank matrices), the original knowledge encoded in the vast majority of parameters is preserved by design. The adaptation is additive (W = W0 + ΔW) and localized, minimizing interference. This makes PEFT not just compute-efficient but also knowledge-preserving, enabling sequential adaptation for multiple downstream tasks without erasing the foundational model's capabilities.

Continual learning is the machine learning paradigm where a model learns from a non-stationary stream of data, acquiring new knowledge over time while retaining previous skills. It is the primary research field dedicated to overcoming catastrophic forgetting.

• Core Challenge: Balancing stability (retaining old knowledge) with plasticity (learning new information).
• Approaches: Include rehearsal (replaying old data), regularization (penalizing changes to important weights), and architectural methods (adding new parameters for new tasks).
• Example: A robot learning to navigate new rooms in a building without forgetting the layout of rooms it learned previously.

Elastic Weight Consolidation is a regularization-based continual learning algorithm that mitigates catastrophic forgetting by identifying and protecting parameters important for previous tasks.

• Mechanism: It calculates a Fisher information matrix to estimate each parameter's importance to learned tasks. During new task training, changes to important parameters are heavily penalized.
• Analogy: Treats the parameter space like an elastic sheet, making important weights "stiff" and less important ones "plastic."
• Contrast with PEFT: While EWC modifies all weights with constraints, PEFT methods like LoRA avoid the issue by freezing the base model and adding isolated, task-specific adapters.

Multi-task learning is a training paradigm where a single model is trained simultaneously on multiple related tasks, sharing representations to improve generalization and data efficiency.

• Relation to Forgetting: It is often presented as an alternative to sequential (continual) learning, as learning tasks concurrently avoids the risk of catastrophic forgetting altogether.
• Limitation: Requires access to all task datasets upfront, which is often impractical in real-world, evolving scenarios.
• PEFT Connection: Methods like LoRA can be used for efficient multi-task learning by training separate, small adapter sets for each task on a shared frozen backbone.

The stability-plasticity dilemma is a fundamental trade-off in adaptive systems, including neural networks, between the ability to retain existing knowledge (stability) and the ability to integrate new information (plasticity).

• Catastrophic Forgetting is the extreme failure case where excessive plasticity destroys stability.
• Biological Basis: Observed in neural systems; the brain must be plastic to learn but stable to remember.
• Engineering Goal: The design of continual learning algorithms, including parameter-efficient fine-tuning (PEFT), is explicitly aimed at navigating this dilemma. PEFT methods like LoRA promote stability by freezing the core model and isolating plasticity to small, modular components.

Task interference is the phenomenon where learning a new task degrades performance on a previously learned task due to overlapping or conflicting updates to shared model parameters. It is the direct mechanistic cause of catastrophic forgetting.

• Mechanism: Gradient updates for Task B overwrite the weight configurations that were optimal for Task A.
• PEFT Mitigation: Techniques like LoRA and other adapter methods reduce interference by constraining updates to a small, low-dimensional subspace (the adapter weights), leaving the vast majority of the base model's parameters untouched and preserving their original knowledge.
• Measurement: Often quantified by the drop in accuracy on Task A after training on Task B.

Gradient Episodic Memory is a rehearsal-based continual learning method that stores a subset of past training examples in a memory buffer and uses them to constrain gradient updates during new task training.

• Core Idea: When computing the gradient for a new task, GEM projects it to a direction that does not increase the loss on examples from the episodic memory of old tasks.
• Advantage: Provides a more direct and theoretically grounded constraint than simple replay, actively preventing catastrophic forgetting.
• Comparison: Unlike pure parameter-efficient methods, GEM requires storing raw data (raising privacy concerns) and additional compute for the projection step, but it can be very effective.