Continuous Learning in Single-Incremental-Task Scenarios

By Davide Maltoni, Vincenzo Lomonaco. 2018

337 citations recorded by the source corpus.

It was recently shown that architectural, regularization and rehearsal strategies can be used to train deep models sequentially on a number of disjoint tasks without forgetting previously acquired knowledge. However, these strategies are still unsatisfactory if the tasks are not disjoint but constitute a single incremental task (e.g., class-incremental learning). In this paper we point out the differences between multi-task and single-incremental-task scenarios and show that well-known approaches such as LWF, EWC and SI are not ideal for incremental task scenarios. A new approach, denoted as AR1, combining architectural and regularization strategies is then specifically proposed. AR1 overhead (in terms of memory and computation) is very small thus making it suitable for online learning. When tested on CORe50 and iCIFAR-100, AR1 outperformed existing regularization strategies by a good margin.

Subfields: Artificial Intelligence, Computational Mechanics, Meta & Continual 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 Computational Mechanics

In Meta & Continual Learning

Citations in the corpus

Re-evaluating Continual Learning Scenarios: A Categorization and Case for Strong Baselines (2018, background): “Such single-headed Split MNIST is known as incremental class learning [11, 12, 13].”
DisCoRL: Continual Reinforcement Learning via Policy Distillation (2019, methodology): “…, either by constraining weight updates in order to maintain knowledge from previous tasks [Kirkpatrick et al., 2017, Zenke et al., 2017, Maltoni and Lomonaco, 2018], or by keeping an old model in memory and distilling knowledge [Hinton et al., 2015] into it later to remember [Li and…”
Task-Aware Information Routing from Common Representation Space in Lifelong Learning (2023, methodology): “Following Maltoni & Lomonaco (2019), Seq-Core50 is organized into nine tasks, the first of which includes ten classes, while the rest have five classes each.”
CODA-Prompt: COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning (2022, background): “One group of approaches expands a model’s architecture as new tasks are encountered; these are highly effective for applications where a model growing with tasks is practical [16, 35, 39, 43, 54].”
Online Continual Learning Via Candidates Voting (2021, methodology): “Instead of using a singlehead classifier [31] for all classes seen so far in classincremental setting, methods for task-incremental problem apply a multi-head classifier [1] for each independent task and domain-incremental methods aim to learn the label shift rather than new classes.”
Total Recall: a Customized Continual Learning Method for Neural Semantic Parsers (2021, background): “ods (Mallya and Lazebnik, 2018; Serra et al., 2018; Maltoni and Lomonaco, 2019; Houlsby et al., 2019; Wang et al., 2020b; Pfeiffer et al., 2021; Rusu et al., 2016) which dynamically alter the structures of models to reduce the catastrophic forgetting; iii)”
Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learning (2021, methodology): “For example, methods which expand a model’s architecture as new tasks are encountered are highly effective for applications where a model growing with tasks is practical [12, 32, 36, 40, 48, 51].”
Continual Learning for Natural Language Generation in Task-oriented Dialog Systems (2020, methodology): “Methods using dynamic architectures (Rusu et al., 2016; Maltoni and Lomonaco, 2019) increase model parameters throughout the continual learning process, which leads to an unfair comparison with other methods.”
Imbalanced Continual Learning with Partitioning Reservoir Sampling (2020, background): “2], and fix the memory size to 2000 (as done in [60]), which are 5.”
Learning to Continually Learn (2020, methodology): “…of contrastive excitation backpropagation [20, 46], by interleaving pseudo examples generated from the current weights to limit plasticity [25], by employing L2 regularization on the weight changes between tasks [24], or by combining regularization and the Copy Weight with Reinit strategy [27, 30].”
Continual Reinforcement Learning in 3D Non-stationary Environments (2019, methodology): “CRL-Naive: This approach, like the homonym strategy in the supervised context [26, 30], consists in just continuing the learning process without variations and indifferently w.”
Continual Learning for Robotics (2019, methodology): “Supervised Continual Learning approaches have been mostly focused on classification (Lopez-Paz and Ranzato, 2017; Maltoni and Lomonaco, 2018; Kirkpatrick et al., 2017).”
Don't forget, there is more than forgetting: new metrics for Continual Learning (2018, methodology): “This cumulative method is a sort of upper bound, or ideal performance that CL strategies should try to reach (Maltoni and Lomonaco, 2018).”