Standard deep learning models operate under a comfortable assumption: all training data is available simultaneously, shuffled into an independent and identically distributed (i.i.d.) stream. However, when a neural network is forced to learn tasks sequentially—such as a drone adapting to changing lighting conditions or a medical model learning from a succession of hospitals—this assumption collapses.
When a network encounters a new data distribution, gradient descent updates its parameters to minimize the error on the current task. Because the loss landscapes of previous tasks are completely hidden from the current optimization step, these updates blindly overwrite the parameter configurations the network spent thousands of steps perfecting. This phenomenon is known as catastrophic forgetting. The network exhibits high plasticity (the ability to absorb new information) but lacks stability (the ability to retain past knowledge).
To build systems capable of continuous adaptation, researchers have developed diverse strategies to balance this stability-plasticity trade-off. Rather than simply blocking parameter updates, modern approaches focus on preserving the geometric structure of representations, simulating past experiences, isolating parameter pathways, or treating learning as a continuous probabilistic fusion.
The Geometry of Representation Space
To understand why forgetting happens, we can look at how a network structures its internal embedding space. When a model learns a task, it maps inputs into clusters or geometric configurations that separate different classes or concepts. Naive sequential training warps this space; as the network accommodates new tasks, the geometric relationships that defined older tasks collapse.
Interestingly, sheer scale provides a partial defense. Ramasesh et al. demonstrate that scaling up both model size and pretraining dataset size systematically increases a network's resistance to forgetting. This resilience emerges because larger, pretrained models naturally develop more orthogonal internal representations. When tasks occupy distinct, non-interfering subspaces within the network's high-dimensional activations, updates for a new task are less likely to collide with the parameters governing an old one.
Standard Network (Low Capacity):
Task 1 Representation ───► [ Shared Weights ] ◄─── Task 2 Updates (Overwrites Task 1)
Large/Pretrained Network (High Capacity):
Task 1 Representation ───► [ Subspace A ] ┐
├─► Non-interfering Orthogonal Updates
Task 2 Representation ───► [ Subspace B ] ┘
When scale alone is insufficient, researchers explicitly anchor the geometry of the embedding space. For example, Pomponi et al. introduce Centroids Matching, an approach that represents learned classes as static centroids within an embedding space. Instead of modifying the entire network architecture, the optimization objective ensures that new learning cycles do not distort the relative geometric distances between previously established feature clusters.
Taking this geometric preservation further, Gao et al. propose expanding the underlying representation space into a mixed-curvature space that combines hyperbolic, spherical, and Euclidean geometries. By dynamically adjusting the space's curvature to match the evolving topology of an incoming data stream, the model preserves complex hierarchical and cyclical relationships between data points, preventing the representational drift that causes forgetting.
Rehearsal and Generative Reminders
The most intuitive solution to forgetting is to remind the network of its past. If a model can periodically revisit old examples while learning new ones, the optimization process can find a joint solution that satisfies both old and new objectives. However, storing raw historical data is often impossible due to strict hardware memory constraints or patient privacy regulations.
To circumvent this, researchers use generative replay to synthesize artificial memories on the fly. In facial expression recognition, de Souza Britto Jr. et al. employ Generative Adversarial Networks (GANs) to generate synthetic images of expressions learned in past datasets, interleaving them with new data. Similarly, in text generation, Xia develops a framework that generates synthetic "review" sentences to preserve named entity recognition capabilities without storing original user text.
When generative models are too expensive to deploy, compressed or partial data storage offers an alternative. Zhai et al. demonstrate that instead of saving full images, storing a tiny fraction of random image patches and reconstructing them using a bilateral Masked Autoencoder (MAE) provides a highly efficient rehearsal mechanism.
Generative Replay:
[Past Task Data] ──► Train GAN ──► [Synthetic Memories] ┐
├─► Joint Training ──► Updated Model
[New Task Data] ─────┘
Compressed Rehearsal (MAE):
[Past Task Data] ──► Mask Patches (Save 20%) ──► MAE Reconstruction ──► Rehearsal Stream
The efficiency of rehearsal also depends on when and how hard a model reviews. Inspired by human psychology, Kang introduces the View-Batch Model, which applies Ebbinghaus' "forgetting curve" theory to neural networks. The model introduces a spacing effect—resting between exposures to the same data points—while utilizing self-supervised learning to force the network to "study harder" during active training windows. To manage the stream dynamically, Kong introduces a trust-region approach that uses confidence scores to sense when a model has mastered a new task, adaptively scaling down rehearsal frequencies to optimize computational overhead.
Parameter Isolation and Modular Tuning
Instead of forcing a shared set of weights to accommodate every task, parameter isolation strategies allocate dedicated neural real estate to different objectives. By freezing critical pathways or adding modular components, new knowledge is routed around existing structures, guaranteeing that old memories remain untouched.
This isolation can happen at various levels of granularity:
- Bit-Level Isolation: Shi et al. operate at the level of individual bits within quantized weights. Their framework, Bit-Level Information Preserving (BLIP), calculates the information gain of specific bits through a Bayesian lens and freezes the most critical bits, preventing subsequent tasks from altering them.
- Activation-Level Masking: Rather than masking weights, Masana et al. introduce Ternary Feature Masks (TFM), which apply masks directly to feature activations. This creates isolated, task-specific pathways through a shared backbone network, achieving a zero-forgetting guarantee with minimal memory overhead.
- Parameter-Efficient Tuning (PET): For massive foundation models, researchers append small, trainable modules while keeping the core network frozen. Wu et al. propose SD-LoRA, which decouples the magnitude and direction of Low-Rank Adaptation updates. By reusing previously learned parameter directions and only tuning their scaling magnitudes, the model absorbs new classes without extra inference costs. Similarly, Gao et al. introduce the Learning-Accumulation-Ensemble (LAE) framework, which calibrates the adaptation speeds of adapters and LoRA layers across a dual-expert ensemble to smoothly integrate sequential tasks.
Weight Sharing (Forgetting):
New Task ──► [ Shared Weights (Overwritten) ]
Ternary Feature Masks (Isolation):
New Task ──► [ Shared Weights ] ──► [ Task-Specific Activation Mask ] ──► Isolated Output
Probabilistic Fusion and Continuous Consolidation
A more fundamental shift involves moving away from deterministic point-estimate weight updates entirely. If we treat a neural network as a probabilistic system, sequential learning becomes a problem of continuous information fusion—updating a distribution over parameters as new evidence arrives.
Li et al. implement this by using Bayesian Neural Networks to maintain a mixture of Gaussian distributions across the parameter space. When a new task is introduced, the model updates its posterior distribution without erasing the prior knowledge distributions associated with earlier tasks.
To bridge the gap between deep networks and exact probabilistic models, Scannell et al. introduce Sparse Function-space Representation (SFR). SFR mathematically converts trained neural networks into sparse Gaussian Processes. This transformation allows the network to inherit the principled uncertainty and sequential data-fusion capabilities of Gaussian Processes while retaining the scalability of deep architectures.
Taking this concept to its logical conclusion, Nabarro introduces GBP Learning, which completely reimagines deep neural networks as Gaussian factor graphs. In this framework, inputs, weights, and activations are all treated as random variables. By utilizing Gaussian Belief Propagation (GBP), the network performs local, asynchronous updates. Learning becomes a continuous, fluid integration of old and new information, eliminating the need for discrete task boundaries or external replay buffers.
On the architectural side, consolidation can be achieved through multi-timescale dynamics. Kaplanis et al. propose Policy Consolidation (PC) for reinforcement learning agents. PC uses a cascade of neural networks that automatically pass knowledge down a chain of timescales. Fast-changing networks absorb immediate environmental transitions, while slower networks gradually consolidate stable, long-term policies, mimicking the memory consolidation processes observed in biological brains.
Key Foundational Work
ASP: Learning to Forget With Adaptive Synaptic Plasticity in Spiking Neural Networks (2017) — contributed 1 concept elements to this direction.