Catastrophic Forgetting

A regularization technique that identifies and protects the synaptic weights most critical to previously learned tasks. It acts as a quadratic penalty, constraining important parameters from deviating far from their old values during new task training.

• Fisher Information Matrix: Calculated to determine the importance of each weight.
• Mechanism: Adds a loss term proportional to the importance-weighted squared distance.
• Benefit: Allows learning new patient-specific patterns without erasing general diagnostic capability.

An architectural approach that instantiates a new neural network column for each new task while retaining the pre-trained columns. Lateral connections from previously frozen columns enable transfer learning without modifying old weights.

• No Forgetting: Old columns are literally frozen and immutable.
• Complexity Cost: Model size grows linearly with the number of tasks.
• Use Case: Ideal for adding new disease detection capabilities to a deployed device without risking prior FDA-cleared functionality.

A data-centric strategy where a small buffer of representative samples from previous tasks is stored locally on the device. During new task training, these old samples are interleaved with new data.

• Memory Buffer: Requires a small, privacy-compliant storage of pseudo-data or raw features.
• Constraint: Strictly limited by the device's RAM and storage budget.
• Variation: Generative Replay uses a separate generative model to synthesize old-task data instead of storing it.

An unsupervised regularization method that computes the importance of each synapse online during the training trajectory. Unlike EWC, it does not require a separate Fisher Information calculation phase.

• Path Integral: Tracks the contribution of each weight change to the total loss reduction.
• Efficiency: Lower computational overhead than EWC, suitable for continuous on-device adaptation.
• Mechanism: Decelerates learning on weights deemed highly important for previous tasks.

An importance estimation technique that judges a weight's significance by the sensitivity of the learned function's output to a change in that weight. It computes the gradient of the squared L2 norm of the model's output.

• Unsupervised: Requires no labeled data to compute importance.
• Local Calculation: Can be computed efficiently on the edge device using only unlabeled sensor streams.
• Advantage: Captures the importance of weights for the function learned, not just the loss landscape.

A distillation-based approach where the model's responses on new task data, as predicted by the old model, are used as soft targets. The new model is trained to match these old outputs while also learning the new task.

• No Old Data: Requires absolutely no storage of previous task data.
• Knowledge Distillation: Uses the old model's logits as a supervisory signal.
• Trade-off: Preserves output behavior but may not perfectly retain internal feature representations.