Architectural Methods

This core mechanism assigns distinct, non-overlapping subsets of a model's parameters to different tasks. By isolating task-specific pathways, it completely avoids inter-task interference and catastrophic forgetting by design. Key implementations include:

• Hard Attention to the Task (HAT): Learns binary attention masks over network neurons to gate activation flow per task.
• Supermasks: Identifies sparse, trainable subnetworks within a larger, frozen model for each new task. This approach is highly effective but can lead to linear parameter growth with the number of tasks.

These methods grow the neural architecture to accommodate new knowledge, freezing old parameters to preserve past learning. The canonical example is Progressive Neural Networks, which adds a new column of layers for each task, with lateral connections to previous columns to enable feature transfer. This provides guaranteed stability but results in a model whose size scales directly with the number of tasks, posing challenges for edge deployment where memory is constrained.

A more parameter-efficient form of isolation where the network maintains a large, shared parameter base, but only a sparse subset is activated for any given input or task. Mechanisms include:

• Mixture-of-Experts (MoE): Routes inputs through different, specialized sub-networks (experts) via a gating network.
• Task-Conditioned Routing: Uses task identifiers or learned embeddings to select specific pathways through a monolithic model. This enables high capacity with sub-linear compute growth, a critical consideration for on-device inference.

This principle involves building complex models from reusable, task-specific modules. New tasks are learned by composing or slightly adapting existing modules, or by adding new ones. This facilitates forward transfer (using old modules for new tasks) and simplifies updates. It aligns with software-defined design patterns, making models more interpretable and easier to manage in long-term lifelong learning scenarios on edge fleets.

Automates the discovery of optimal network structures for continual learning. Techniques like Neural Architecture Search (NAS) or continual learning-aware pruning can dynamically identify which parts of a network to expand, freeze, or prune when a new task arrives. This meta-approach aims to balance the stability-plasticity dilemma automatically, optimizing for metrics like final accuracy, memory footprint, and backward transfer.

Most practical systems combine architectural changes with other continual learning strategies. Common hybrids include:

• Expansion + Rehearsal: A dynamically growing network uses a small replay buffer to stabilize learning within new modules.
• Isolation + Regularization: Task-specific parameters are isolated, but a regularization term (like from Elastic Weight Consolidation) is applied within each module to prevent internal forgetting. These hybrids are essential for achieving robust performance in challenging online continual learning settings on edge devices.

An architectural method that freezes a neural column after learning a task and adds new, laterally connected columns for subsequent tasks. This prevents forgetting by design, as old parameters are immutable. However, it leads to linear parameter growth, making it less suitable for long task sequences on edge devices.

• Key Mechanism: Lateral connections from old to new columns allow the new column to leverage previously learned features.
• Primary Use: Task-incremental learning scenarios where computational growth is acceptable.

A method that learns task-specific binary attention masks over network neurons. For each new task, a sparse mask is learned, allowing selective activation of a subnetwork. This enables parameter sharing while isolating task-specific pathways.

• Key Mechanism: A sigmoid-based hard attention mechanism gates neuron activations, controlled by task-specific embedding vectors.
• Primary Use: Class-incremental and domain-incremental learning with a fixed parameter budget.

A family of methods that assign distinct, non-overlapping subsets of model parameters to different tasks. This is the most direct architectural approach to prevent interference, as tasks do not share weights. It includes techniques like PackNet, which iteratively prunes and freezes weights for old tasks before allocating new capacity.

• Key Mechanism: Task-specific parameter allocation via pruning, masking, or expansion.
• Challenge: Requires intelligent capacity budgeting and can be inefficient if task similarity is high.

Methods that grow the network architecture in response to new tasks, adding neurons, layers, or branches. This contrasts with fixed-capacity regularization methods. The expansion can be triggered by novelty detection or task performance plateaus.

• Examples: Dynamically Expandable Networks (DEN) and Progressive Nets.
• Edge Consideration: Uncontrolled growth is prohibitive for edge devices, necessitating growth budgets and selective pruning.

An architecture combining a gating network with an array of task-specific expert models. For a given input, the gating network routes the sample to the most relevant expert. This isolates task knowledge within each expert while the gate learns task relationships.

• Key Mechanism: Mixture-of-Experts (MoE) paradigm adapted for continual learning.
• Primary Use: Task-incremental learning where task identity is inferred at test time.

A compositional approach where a model is constructed from a library of reusable neural modules. New tasks are learned by assembling and fine-tuning a new configuration of these modules, promoting knowledge reuse and minimizing new parameters.

• Key Mechanism: Module selection and composition via reinforcement learning or gradient-based search.
• Benefit: Encourages positive forward transfer by recombining previously useful functional units.