Streaming Continual Learning for Unified Adaptive Intelligence in Dynamic Environments

TL;DR

A unified setting is put forward that harnesses the benefits of both CL and SML: their ability to quickly adapt to nonstationary data streams without forgetting previous knowledge, known as streaming CL (SCL).

Abstract

Developing effective predictive models becomes challenging in dynamic environments that continuously produce data and constantly change. Continual Learning (CL) and Streaming Machine Learning (SML) are two research areas that tackle this arduous task. We put forward a unified setting that harnesses the benefits of both CL and SML: their ability to quickly adapt to non-stationary data streams without forgetting previous knowledge. We refer to this setting as Streaming Continual Learning (SCL). SCL does not replace either CL or SML. Instead, it extends the techniques and approaches considered by both fields. We start by briefly describing CL and SML and unifying the languages of the two frameworks. We then present the key features of SCL. We finally highlight the importance of bridging the two communities to advance the field of intelligent systems.

Figures (2)

• Figure 1: Comparison between a Continual Learning algorithm $A^{\text{CL}}$ and Streaming Machine Learning algorithm $A^{\text{SML}}$ trained on a possibly infinite sequence of items, indexed by the subscript $i$. (\ref{['fig:cl']}) Continual Learning with a model $h$ and an external memory $M$. Training is performed on one training item's data $D_i$ at a time. The test sets $D_1^{\text{test}}, D_2^{\text{test}}, \ldots, D_i^{\text{test}}$ are used for evaluation. (\ref{['fig:sml']}) Streaming Machine Learning with a model $h$ and a stream of items' data $D_i$, where each $D_i$ contains a mini-batch of a few examples. Each $D_i$ is used both for the testing and the training phases.
• Figure 2: Comparison between Continual Learning Scenarios and Drifts in Streaming Machine Learning. The red dashed line highlights the decision boundary, while grey dots represent data no longer available. Different shapes represent novel data that may still fall into the same classes as the previous experience, as per Domain Incremental Learning, or may represent new classes (denoted by different colors), as per Class Incremental Learning. In Domain-Incremental learning, each experience contains examples from all classes, while new experiences introduce new examples only, potentially leading to virtual drift. Class-Incremental learning is prone to concept evolution since it exclusively introduces examples of new classes in each new experience. Task-Incremental learning assigns a task label to each example. It can be viewed as a concept drift or a concept evolution, depending on whether we consider the class labels of all tasks or only those of the current task.