Catastrophic Forgetting

This is the core tension at the heart of catastrophic forgetting. A network must possess plasticity to rapidly encode new information from a fresh manufacturing dataset, but it also requires stability to retain previously learned representations without degradation. When a model is fine-tuned exclusively on new data, the gradient updates that minimize the new task's loss function will unconstrainedly shift the weights critical for the old task, effectively overwriting the established decision boundaries.

Forgetting occurs because of destructive interference in the shared weight space of the neural network. When learning Task B, the optimizer adjusts parameters that were highly sensitive for Task A. This is particularly acute in deep neural networks where representations are highly distributed. A small change in a single layer's weights can cascade through the network, completely corrupting a previously learned feature detector. This is distinct from mere capacity overflow; the network has the theoretical capacity to learn both tasks, but the gradient-based optimization process lacks a native mechanism to protect old knowledge.

A direct strategy to prevent forgetting is to interleave examples from previous tasks during new training. This is known as experience replay or pseudo-rehearsal. Key approaches include:

• Exact Rehearsal: Storing a subset of old manufacturing data and mixing it with new data in each batch.
• Generative Rehearsal: Training a generative model to produce synthetic samples of old data, bypassing the need for a large storage buffer. This method explicitly reinforces old decision boundaries, but the storage of historical data can be problematic in memory-constrained edge devices or when data retention policies are strict.

These techniques protect important parameters from being drastically altered. Instead of replaying data, they add a penalty term to the loss function that anchors critical weights. Elastic Weight Consolidation (EWC) calculates the Fisher Information Matrix to identify weights that are crucial for previous tasks and penalizes their movement quadratically. Synaptic Intelligence (SI) tracks the contribution of each synapse to the reduction in loss over a task's training trajectory, consolidating those pathways. This approach is highly parameter-efficient, as it doesn't require storing old data, making it suitable for continuous learning on a factory-floor inference node.

A structural solution involves allocating distinct, non-overlapping parameters to different tasks to eliminate interference entirely. This can be achieved through:

• Progressive Neural Networks: Freezing the original network and adding new lateral columns for each new task.
• Dynamic Sparse Training: Training a sparse sub-network for each task within a single, large fixed model, activating only the relevant pathways at inference time. While extremely robust against forgetting, these methods introduce a linear growth in parameters with the number of tasks, which can be a scalability challenge for a foundation model expected to master hundreds of distinct manufacturing processes.

This approach uses the model's own previous state as a teacher to preserve its behavior. When learning a new task, the model's output on the new data is regularized to not diverge from the output of a frozen copy of the model from the previous task. Learning without Forgetting (LwF) is a seminal technique where the old model's softened logits on the new task's data serve as a distillation target, encouraging the new model to maintain the old feature extraction patterns even when old data is completely unavailable. This is particularly useful when historical manufacturing data cannot be retained for compliance reasons.

A regularization technique that identifies and protects the synaptic weights most critical for previously learned tasks. It works by calculating the Fisher information matrix to measure the importance of each parameter. When learning a new task, EWC applies a quadratic penalty that constrains changes to these high-importance weights, forcing the model to encode new knowledge in less critical parameters. This mimics the synaptic consolidation observed in biological brains, allowing sequential learning without catastrophic interference.

A biologically inspired method that interleaves samples from previous tasks into the training stream for new tasks. By maintaining a memory buffer of representative examples, the model continuously rehearses old knowledge. Key implementation variants include:

• Reservoir sampling: Randomly populating a fixed-size buffer
• Maximally Interfered Retrieval: Selecting replay samples that are most likely to be forgotten
• Generative Replay: Training a separate generative model to produce synthetic samples of old tasks instead of storing raw data

An architectural solution that sidesteps forgetting entirely by instantiating a new neural network column for each new task. Lateral connections from previously frozen columns allow the new column to leverage learned features without modifying them. While this guarantees zero forgetting, the parameter count grows linearly with the number of tasks. This approach is particularly relevant for industrial robotics where a sequence of distinct manipulation skills must be accumulated without degradation.

A distillation-based approach where the model's responses on the new task are constrained to remain consistent with its previous outputs on old tasks. Before training on new data, the model records its own soft targets for old task examples. During optimization, a distillation loss encourages the updated model to match these recorded outputs. This requires only the new task data and the model's own recorded responses, eliminating the need for a raw data buffer.

Quantifying catastrophic forgetting requires specialized metrics beyond standard accuracy:

• Backward Transfer (BWT): Measures how learning a new task impacts performance on previously learned tasks. A negative BWT indicates forgetting.
• Forward Transfer (FWT): Measures how learning previous tasks improves zero-shot performance on future tasks.
• Average Accuracy: The mean performance across all tasks after sequential training.
• Forgetting Measure: The difference between a task's peak performance and its final performance after all subsequent learning.