Source-linked AI summary
Learning to Remember: A Synaptic Plasticity Driven Framework for Continual Learning
Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jähnichen, Moin Nabi
TL;DR
Continual learning must preserve prior knowledge while learning new classes from a stream without storing old real samples and while scaling capacity over time. The paper introduces Dynamic Generative Memory, which uses a conditional generative adversarial network with learnable neural masks and adaptive expansion. In visual class-incremental experiments, DGM outperforms cited generative-memory methods, while weight masking is more efficient than activation masking.
Problem
Continual learning must retain knowledge across class-incremental tasks without replaying stored real samples, while accommodating continuously growing information.
Method
DGM uses a single conditional generative adversarial memory with learned masks on weights or activations and expands the network by the number of parameters blocked for the previous task.
Results
DGM outperforms cited state-of-the-art generative-memory methods on MNIST and SVHN, and DGMw is more efficient than DGMa.
Takeaways & Limitations
Generative memory with neural masking and dynamic expansion supports strict single-head class-incremental learning without storing old real samples.
Takeaways & Limitations
The authors identify missing backward knowledge transfer and limited saturation of the network growth pattern as limitations for future work.
Abstract
from arXiv · showhide
Models trained in the context of continual learning (CL) should be able to learn from a stream of data over an undefined period of time. The main challenges herein are: 1) maintaining old knowledge while simultaneously benefiting from it when learning new tasks, and 2) guaranteeing model scalability with a growing amount of data to learn from. In order to tackle these challenges, we introduce Dynamic Generative Memory (DGM) - a synaptic plasticity driven framework for continual learning. DGM relies on conditional generative adversarial networks with learnable connection plasticity realized with neural masking. Specifically, we evaluate two variants of neural masking: applied to (i) layer activations and (ii) to connection weights directly. Furthermore, we propose a dynamic network expansion mechanism that ensures sufficient model capacity to accommodate for continually incoming tasks. The amount of added capacity is determined dynamically from the learned binary mask. We evaluate DGM in the continual class-incremental setup on visual classification tasks.
1. Introduction
The paper targets strict class-incremental learning, where models must preserve prior knowledge without storing old real samples while accommodating continually growing information. It introduces DGM, which combines generative memory, learnable neural masking, and adaptive network expansion.
- Continual learning must prevent catastrophic forgetting, scale with growing data, and transfer knowledge across tasks.
- Strict class-incremental learning evaluates all observed classes without task labels and forbids access to previously seen real samples.Generative memory instead replays samples synthesized from memorized previous data distributions.
- DGM uses a single generator with learnable connection plasticity represented by parameter-level attention masks.The framework learns masks for generator weights or activations during adversarial training and incrementally learns new tasks without replaying previous knowledge to the generator.
- DGM dynamically expands the network by adding exactly the parameters blocked for the previous task, keeping the number of freely trainable parameters constant.This mechanism is designed to maintain sufficient capacity and constant expressive power as tasks accumulate.
- The paper evaluates DGMw, using weight masks, and DGMa, using activation masks, for continual visual classification.
2. Related Work
Prior continual-learning methods address forgetting through parameter protection, task-specific subspaces, exemplar replay, or generative replay. DGM differs by combining hard attention in a single generative memory with dynamically expanded capacity.
- Parameter-based methods: Earlier synaptic-plasticity approaches primarily target task-incremental learning, where inference uses an oracle task label and a dedicated output head.
- Replay-based methods: Class-incremental methods such as iCarl retain representative real samples, but raw-data storage conflicts with privacy and memory constraints.
- Generative memory: Generative replay avoids storing old samples but retrains generators on synthesized previous data and can suffer from inefficient training and semantic drift.
- Generative memory: DGM uses hard attention in a generator to avoid replaying previous knowledge to that generator while learning new tasks.
- Network expansion: DGM expands generator capacity dynamically as knowledge grows, whereas a related approach duplicates neurons subject to sharp semantic changes.
- Memory organization: Unlike approaches with separate short- and long-term networks, DGM embeds memory transfer in one network through binary mask learning.
3. Dynamic Generative Memory
DGM combines conditional generative memory with binary neural masks to learn successive tasks while preserving previously used parameters. It dynamically expands network capacity to keep free capacity constant during continual class-incremental training.
- 3.1. Learning Binary Masks: Task-specific binary masks are learned for generator weights or layer activations and combined with layer weights during computation.DGMw uses masks shaped like weight matrices; DGMa expands neuron masks across the corresponding weight matrix.
- 3.1. Learning Binary Masks: The mask scaling parameter is annealed during training, while DGMa uses only global annealing because it showed better performance.DGMw uses incremental annealing during epochs; the global schedule varies across epochs.
- 3.1. Learning Binary Masks: Gradients are multiplied by the reverse cumulative mask to prevent overwriting generator knowledge associated with previous classes.A sparsity regularizer also promotes reuse of previously reserved units rather than reserving new ones.
- 3.2. Dynamic Network Expansion: Dynamic expansion restores the capacity reserved for each task, keeping the number of free neurons or weights constant across learning cycles.For DGMa, output neurons increase by the number reserved; for DGMw, output neurons expand by approximately the reserved weights divided by input size.
- 3.3. Training of DGM: DGM jointly trains generative, discriminative, and classification objectives in a strictly class-incremental setup.The generator synthesizes task-conditioned images, while the discriminator also serves as the classifier.
- 3.3. Training of DGM: The generator optimizes classification and adversarial losses with regularization weighted by the ratio of used to free network capacity.This weighting encourages less reuse early and more reuse later as the model becomes occupied.
4. Experimental Results
DGM is evaluated on strict class-incremental visual classification benchmarks and compared with generative-replay and episodic-memory baselines. Results examine classification, generated samples, mask evolution, network growth, and storage efficiency.
- Benchmark setup: DGM is evaluated on MNIST, SVHN, CIFAR-10, and ImageNet-50 using average accuracy over all classes seen so far.The experiments use a strictly class-incremental setup, with additional comparisons allowing partial storage of real samples.
- Classification performance: DGM outperforms state-of-the-art generative-memory methods on MNIST and SVHN, while DGMw improves over MeRGAN’s replay-based approach.The reported comparison attributes the gains to integrating memory learning into the generator and expanding the network as it saturates.
- Generated samples: DGM generates informative, diverse, and class-discriminative samples after incremental training on MNIST, SVHN, and ImageNet.The ImageNet evaluation adds 10 classes per step and observes discriminative features that remain after five tasks.
- Classification performance: Adding real samples to replay improves DGM beyond iCarl, although DGM without real samples is sometimes outperformed at later class counts.With equal memory size, DGM outperforms iCarl in top-1 accuracy after 30 classes and nearly matches it in top-5 accuracy; fixed memory reduces replayed samples per class over time.
- Plasticity and memory usage: Mask learning shifts from short-term, partially binary allocations toward long-term reservations, while later tasks reuse more units and block fewer new ones.The reuse trend is associated with more efficient capacity allocation; generator storage for ImageNet-50 is 228MB versus 315MB for preprocessed training samples, a 27.5% reduction.
5. Conclusion
DGM addresses single-head, strictly incremental class-incremental learning by using a conditional generative adversarial model whose generator serves as a masked memory module. The authors report more efficient weight masking than activation masking, while identifying missing backward knowledge transfer and limited network-growth saturation as limitations.
- DGM targets single-head, strictly incremental class-incremental learning with a conditional generative adversarial model.
- The generator functions as a memory module endowed with neural masking to address catastrophic forgetting.
- Neural masking is more efficient when applied directly to layer weights rather than activations.
- The authors identify missing backward knowledge transfer and limited saturation of the network growth pattern as DGM limitations.