Abstract
The human brain is the gold standard of adaptive learning. It not only can learn and benefit from experience, but also can adapt to new situations. In contrast, deep neural networks only learn one sophisticated but fixed mapping from inputs to outputs. This limits their applicability to more dynamic situations, where the input to output mapping may change with different contexts. A salient example is continual learning—learning new independent tasks sequentially without forgetting previous tasks. Continual learning of multiple tasks in artificial neural networks using gradient descent leads to catastrophic forgetting, whereby a previously learned mapping of an old task is erased when learning new mappings for new tasks. Herein, we propose a new biologically plausible type of deep neural network with extra, out-of-network, task-dependent biasing units to accommodate these dynamic situations. This allows, for the first time, a single network to learn potentially unlimited parallel input to output mappings, and to switch on the fly between them at runtime. Biasing units are programmed by leveraging beneficial perturbations (opposite to well-known adversarial perturbations) for each task. Beneficial perturbations for a given task bias the network toward that task, essentially switching the network into a different mode to process that task. This largely eliminates catastrophic interference between tasks. Our approach is memory-efficient and parameter-efficient, can accommodate many tasks, and achieves the state-of-the-art performance across different tasks and domains.
Introduction: The Challenge of Catastrophic Forgetting
Deep neural networks are notoriously rigid. When a model is trained on Task A and then moves to Task B, the gradient descent process typically overwrites the weights needed for Task A—a phenomenon known as catastrophic forgetting. Existing solutions generally fall into four categories: weight constraints (e.g., EWC), network expansion, episodic memory replay, or architectural partitioning. However, these methods often face trade-offs between parameter growth, memory intensity, and finite capacity. This research introduces a "Type 5" approach: a biologically plausible method that is both parameter-efficient and capable of handling a nearly unlimited sequence of tasks.
Methodology: The Beneficial Perturbation Network (BPN)
The core innovation is the Beneficial Perturbation Network (BPN). Instead of seeking a single set of weights that satisfies every task simultaneously, BPN utilizes a shared "backbone" of weights and adds out-of-network, task-dependent bias units.
Beneficial Perturbations as a "Mode Switch"
While adversarial perturbations are designed to make a network fail, BPN flips this logic. It calculates "beneficial" noise that pushes the network's internal activations back toward the optimal working region for a specific task, even if the shared weights have drifted due to subsequent learning. These bias units are added to every layer and, during training, are factorized into small matrices to maintain a minimal footprint. Once a task is learned, these are collapsed into a single vector and frozen. BPN is often paired with constraint methods like Elastic Weight Consolidation (EWC) to preserve the shared backbone while the bias units provide per-task corrections.
Key Insights: Flipping the Adversarial Script
- Activation Space Targeting: Unlike adversarial attacks that modify input pixels, BPN modifies internal activations. This targets high-level features already learned by the network, making the process highly efficient.
- Biological Inspiration: The architecture draws parallels to the human brain's hippocampus, which is thought to store task-specific "indexes" that switch the neocortex into different processing modes.
- Minimal Overhead: The bias units introduce negligible overhead—only about a 0.3% increase in parameters per task compared to the 100% increase required for training separate networks.
Results and Comparison
The authors evaluated BPN against standard benchmarks including sequential MNIST, CIFAR-100, and a challenging 8-dataset sequential object recognition task.
Key findings include:
- Superior Accuracy: On the 8-dataset recognition task, BPN (BD+PSP) achieved 58.87% average accuracy, significantly outperforming standard PSP (54.75%) and other methods like LwF.
- Massive Capacity: In a 100-task permuted MNIST test, BPN maintained high performance long after other methods degraded, achieving 30.14% better performance than the original PSP method.
- Memory Efficiency: BPN's memory footprint was found to be up to 300x smaller than episodic memory-based competitors because it only stores small bias vectors rather than raw data.
Conclusion and Future Work
BPN represents a significant step toward general adaptive AI by allowing a single network to switch between unlimited tasks with minimal parameter growth. While currently requiring a "task oracle" to activate the correct bias units at test time, future work will explore automatic context detection and the scalability of beneficial perturbations in deeper architectures like Transformers.
Summary updated