Catastrophic Interference

In a standard multi-layer perceptron, knowledge is distributed across many connection weights rather than localized to specific neurons. Learning a new task involves adjusting these same weights via gradient descent. Since the optimization process has no inherent mechanism to protect weights crucial for prior tasks, the representational overlap means updates for Task B inevitably overwrite the weight configurations that encoded Task A. This is the primary architectural cause of interference.

The sequential nature of SGD-based training is a direct cause. When presented with a stream of data from Task A followed by Task B, the optimizer receives gradients computed solely on the current mini-batch (Task B). These gradients point in a direction that minimizes loss for the new data, with no constraint to preserve the loss on old data. The parameter update $\theta_{t+1} = \theta_t - \eta \nabla L_B(\theta_t)$ moves the network into a configuration optimal for B but potentially catastrophic for A.

Interference severity depends on the similarity of the internal representations required for different tasks. If two tasks (e.g., classifying cats vs. dogs) use highly overlapping features, learning the second may refine rather than destroy. However, for disparate tasks (e.g., language translation followed by image segmentation), the network must repurpose the same limited representational capacity, leading to destructive overwriting. Networks with insufficient capacity suffer more acutely.

This is the fundamental trade-off at the heart of continual learning. Plasticity is the network's ability to learn new information (adapt to Task B). Stability is its ability to retain old information (remember Task A). Standard neural networks are highly plastic but lack stability. Catastrophic interference occurs when optimization maximizes plasticity for the current task at the direct expense of stability for all prior tasks. Mitigation techniques like elastic weight consolidation explicitly manage this trade-off.

It's critical to distinguish general forgetting from catastrophic interference. Forgetting in biological systems is often gradual due to decay. In artificial neural networks, interference is active and abrupt: the very act of learning new patterns (via backpropagation) directly corrupts the weights for old patterns. The loss on Task A doesn't just plateau—it can spike dramatically within a few training iterations on Task B, hence the term 'catastrophic.'

Catastrophic interference is a problem of non-IID (non-independent and identically distributed) sequential data. In standard batch training, data from all tasks is shuffled together (IID assumption), allowing gradients to average across tasks and find a joint solution. In sequential or online learning, the data distribution shifts abruptly. The optimizer, seeing only the new distribution, converges to a local minimum for the current task, which is often far from the minimum for previous tasks.

Elastic Weight Consolidation is a regularization-based continual learning method that estimates the importance of each neural network parameter for previously learned tasks. It then applies a quadratic penalty that slows down learning on important weights when training on new data.

• Core Mechanism: Approximates the Fisher information matrix to measure parameter importance.
• Analogy: Acts like a "spring" attached to important weights, providing resistance to change.
• Key Benefit: Enables sequential task learning without explicit access to old data, though performance depends on the accuracy of the importance estimation.

Gradient Episodic Memory is a continual learning algorithm that stores a small subset of raw training examples from previous tasks in an episodic memory buffer. During training on a new task, it projects the proposed gradient update to ensure it does not increase the loss on the stored memory examples.

• Core Mechanism: Solves a quadratic programming problem to find an update direction that satisfies constraints from past tasks.
• Guarantee: Theoretically ensures non-increasing loss on previous tasks.
• Practical Consideration: The computational cost of the projection step scales with the size of the episodic memory.

Generative Replay is a rehearsal-based technique where a generative model (e.g., a Generative Adversarial Network or Variational Autoencoder) is trained to produce synthetic data samples that mimic the distribution of previous tasks. These generated samples are interleaved with new task data during training.

• Core Mechanism: Uses a separate generative model as a "pseudo-rehearsal" system to simulate past experiences.
• Key Challenge: Requires training and maintaining a high-fidelity generative model that does not itself suffer from catastrophic forgetting.
• Variation: Often paired with a dual-model architecture where one model learns the task and another learns to generate data.

Online Learning is a machine learning paradigm where a model learns sequentially from a continuous stream of data, updating its parameters incrementally with each new data point or small batch. This is in contrast to batch learning, which assumes access to the entire dataset at once.

• Core Challenge: Directly exposed to catastrophic interference due to sequential, non-i.i.d. data.
• System Design: Requires architectures for incremental updates, concept drift detection, and feedback loops.
• Applications: Essential for real-time recommendation systems, algorithmic trading, and adaptive control systems.

The Stability-Plasticity Dilemma is a fundamental trade-off in adaptive systems, including neural networks. Stability refers to the ability to retain previously learned knowledge, while Plasticity is the ability to integrate new information.

• Catastrophic Forgetting: Represents a failure mode where excessive plasticity destroys stability.
• Optimal Balance: The goal of continual learning algorithms is to find an optimal operating point on this spectrum.
• Biological Analogy: Mirrored in neurobiological systems, which must be stable enough to maintain long-term memories yet plastic enough to learn from new experiences.

Multi-Task Learning is a paradigm where a single model is trained simultaneously on multiple related tasks, leveraging shared representations to improve generalization and performance on all tasks. It is often contrasted with sequential (continual) task learning.

• Key Difference: Assumes joint access to all tasks' data during a single training phase, avoiding the sequential interference problem.
• Shared vs. Task-Specific Parameters: Often uses architectures with a shared backbone and task-specific heads.
• Relation to Catastrophic Interference: Multi-task learning provides an upper-bound performance baseline that continual learning algorithms strive to approach without the benefit of simultaneous data access.