Continual Learning in Practice

By Tom Diethe, Tom Borchert, Eno Thereska, Borja Balle, Neil D. Lawrence. 2019

78 citations recorded by the source corpus.

This paper describes a reference architecture for self-maintaining systems that can learn continually, as data arrives. In environments where data evolves, we need architectures that manage Machine Learning (ML) models in production, adapt to shifting data distributions, cope with outliers, retrain when necessary, and adapt to new tasks. This represents continual AutoML or Automatically Adaptive Machine Learning. We describe the challenges and proposes a reference architecture.

Subfields: 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 Meta & Continual Learning

Citations in the corpus

Online Continual Learning in Image Classification: An Empirical Survey (2021, background): “[150] presents a reference architecture for self-maintaining intelligent systems that can adapt to shifting data distributions, cope with outliers, retrain when necessary, and learn new tasks incrementally.”
Meta-Consolidation for Continual Learning (2020, background): “Another kind of methods [51, 72, 69, 19, 68, 62, 87] attempt to expand the network to increase the capacity of the model, while learning new tasks.”
Optimal Continual Learning has Perfect Memory and is NP-hard (2020, background): “CL algorithms are not only a challenging research topic, they are also of tremendous practical importance (see e.g. Diethe et al., 2018): Often, it is impractical or even impos-1 sible to re-train a model every time new data arrives.”
Efficient Architecture Search for Continual Learning (2020, background): “Recently, there are remarkable advances made in many applications based on continual learning: sequential task processing [12], streaming data processing [13] and self-management of resources [1, 14], etc.”
Continual Learning in Deep Networks: an Analysis of the Last Layer (2021, background): “However, even if replay might be necessary for continual learning scenarios [7, 21] , replaying data may have consequences in the memory needs and compute needs.”
Continual Learning, Fast and Slow (2021, background): “Separating the slow learning phase allows DualNets continuously improve representation even when there are no labeled samples, which is ubiquitous in real-world deployment scenarios where labeled data are delayed [21] or even limited, which we will demonstrate in Sec.”
Continual Learning with Foundation Models: An Empirical Study of Latent Replay (2022, background): “This choice is motivated by the generality of the replay strategy as it can be applicable in most CL scenarios (Kalifou et al., 2019; Diethe et al., 2018; Lesort et al., 2019; Prabhu et al., 2020; Traoré et al., 2019; Belouadah and Popescu, 2018; Wu et al., 2019; Hou et al., 2019; Lesort, 2020;…”