Overcoming catastrophic forgetting with hard attention to the task

By Joan Serrà, D'idac Sur'is, Marius Miron, Alexandros Karatzoglou. 2018

1,304 citations recorded by the source corpus.

Catastrophic forgetting occurs when a neural network loses the information learned in a previous task after training on subsequent tasks. This problem remains a hurdle for artificial intelligence systems with sequential learning capabilities. In this paper, we propose a task-based hard attention mechanism that preserves previous tasks' information without affecting the current task's learning. A hard attention mask is learned concurrently to every task, through stochastic gradient descent, and previous masks are exploited to condition such learning. We show that the proposed mechanism is effective for reducing catastrophic forgetting, cutting current rates by 45 to 80%. We also show that it is robust to different hyperparameter choices, and that it offers a number of monitoring capabilities. The approach features the possibility to control both the stability and compactness of the learned knowledge, which we believe makes it also attractive for online learning or network compression applications.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Toward Understanding Catastrophic Forgetting in Continual Learning (2019, background): “In recent years, interest in continual learning has risen [1, 21, 26, 27, 29, 39, 40, 42, 51], especially from the deep learning research community, due to its potential to reduce training time and training set sizes (e.g., by continuously adapting from previous models), both of which are critical…”
Re-evaluating Continual Learning Scenarios: A Categorization and Case for Strong Baselines (2018, background): “) and use one of them in one task [17, 18, 19, 20].”
Continual Lifelong Learning with Neural Networks: A Review (2018, methodology): “an analytical tool to discuss the relatedness and enabling conditions for continual learning in terms of the opportunities to re-use previously acquired functionality in subsequent learning episodes. Serrà et al., (2018) incorporated a hard attention (HAT) mask into the loss function and showed that”
Incorporating neuro-inspired adaptability for continual learning in artificial intelligence (2023, background): “Representative strategies include selectively stabilizing parameters [7–11], recovering old data distributions [12–16] and allocating dedicated parameter subspaces [17, 18], etc.”
Learnability and Algorithm for Continual Learning (2023, methodology): “In training the network hk ◦ fk using the data of task k and the generated pseudo feature vectors, we employ the hard attention mask (Serrà et al., 2018) to prevent forgetting in the feature extractor.”
SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model (2023, background): “…approaches [48, 29, 3, 43, 39, 38], which approximate and recover the previously-learned data distributions; architecture-based approaches [33, 32, 41, 49], which allocate dedicated parameter sub-spaces for each incremental task; etc. Recent works have increasingly explored the benefits…”
A soft nearest-neighbor framework for continual semi-supervised learning (2022, neutral): “Finally, the third category is parameter isolation methods [42, 44], which operate by allocating task-specific parameters.”
A Theoretical Study on Solving Continual Learning (2022, methodology): “HAT [3] and CAT [52] protect previous tasks by masking the important parameters to those tasks.”
Beyond Not-Forgetting: Continual Learning with Backward Knowledge Transfer (2022, methodology): “(1) EWC [12]: a regularization-based method that leverages Fisher Information matrix for weights importance evaluation; (2) HAT [26]: learns a hard attention mask to preserve the knowledge of old task; (3) Orthogonal Weight Modulation (OWM) [32]: learns a projector matrix to project the gradient of ”
Generative Negative Text Replay for Continual Vision-Language Pretraining (2022, methodology): “The third is the structure methods [49,39], which keep old parameters fixed while growing and allocating weights for learning new data.”
On the Effectiveness of Lipschitz-Driven Rehearsal in Continual Learning (2022, methodology): “Differently, structural methods purposefully organize the allocation of model capacity to prevent interference and facilitate parameter sharing [1, 51, 66, 37].”
Continual Learning by Modeling Intra-Class Variation (2022, methodology): “(2022) perform additional steps to freeze or prune the old network parameters to prevent forgetting, while HAT (Serra et al., 2018) utilizes a hard attention mask to constrain the amount of updates of important neurons.”
Class Gradient Projection For Continual Learning (2022, methodology): “Prevalent approaches to address the catastrophic forgetting problem in continual learning can be grouped into three broad categories: regularization-based approaches [2, 22, 30, 45], expansion-based approaches [17, 25, 31, 34, 36, 40], and memory-based approaches [3, 27, 33, 38].”
Long-Tailed Class Incremental Learning (2022, methodology): “Parameter-isolation methods increase model plasticity by adding more neurons, modules [31,33], branches [23] or masks [26,25,34].”
DLCFT: Deep Linear Continual Fine-Tuning for General Incremental Learning (2022, methodology): “Parameter isolation methods [3, 29, 40, 42] are based on dynamically allocating a set of parameters for each task so that the training does not interfere with each other.”
Class-Incremental Learning with Cross-Space Clustering and Controlled Transfer (2022, methodology): “Architecture-based methods modify the network architecture dynamically to fit more tasks, by expanding the model by adding more weights [55, 52], or masking and allocating subnetworks to specific tasks [45], or by gating the parameters dynamically using a task identifier [1].”
S-Prompts Learning with Pre-trained Transformers: An Occam's Razor for Domain Incremental Learning (2022, background): “As presented in the main paper, there exist some expansion-based continual learning methods like [69, 34, 17, 59, 25, 19, 39].”