A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning

TL;DR

The paper presents CN-DPM, an expansion-based approach for task-free continual learning that uses a Dirichlet process mixture of neural experts. Each expert combines discriminative (p(y|x)) and generative (p(x)) components, with online variational inference guiding when to create new experts and STM-based sleep phases to stabilize learning. By modeling the data stream as a mixture and inferring the responsible expert without task labels, CN-DPM achieves strong performance on both discriminative and generative tasks while mitigating catastrophic forgetting. The work demonstrates that adaptive, nonparametric expansion can handle non-stationary, task-agnostic environments and scales to complex splits like Split-CIFAR100.

Abstract

Despite the growing interest in continual learning, most of its contemporary works have been studied in a rather restricted setting where tasks are clearly distinguishable, and task boundaries are known during training. However, if our goal is to develop an algorithm that learns as humans do, this setting is far from realistic, and it is essential to develop a methodology that works in a task-free manner. Meanwhile, among several branches of continual learning, expansion-based methods have the advantage of eliminating catastrophic forgetting by allocating new resources to learn new data. In this work, we propose an expansion-based approach for task-free continual learning. Our model, named Continual Neural Dirichlet Process Mixture (CN-DPM), consists of a set of neural network experts that are in charge of a subset of the data. CN-DPM expands the number of experts in a principled way under the Bayesian nonparametric framework. With extensive experiments, we show that our model successfully performs task-free continual learning for both discriminative and generative tasks such as image classification and image generation.

Figures (10)

• Figure 1: Overview of our CN-DPM model. Each expert $k$ (blue boxes) contains a discriminative component for modeling $p(y|x; \phi_k^D)$ and a generative component for modeling $p(x;\phi_k^G)$, jointly representing $p(x,y;\phi_k)$. We also keep the assigned data count $N_k$ per expert. (a) During training, each sample $(x,y)$ coming in a sequence is evaluated by every expert to calculate the responsibility $\rho_k$ of each expert. If $\rho_{K+1}$ is high enough, i.e., none of the existing experts is responsible, the data is stored into short-term memory (STM). Otherwise, it is learned by the corresponding expert. When STM is full, a new expert is created from the data in STM. (b) Since CN-DPM is a generative model, we first compute the joint distribution $p(x,y)$ for a given $x$, from which it is trivial to infer $p(y|x)$.
• Figure 2: Split-CIFAR10 (0.2 Epoch).
• Figure 3: Split-CIFAR100.
• Figure 4: An example of the expert pruning in the Split-MNIST scenario.
• Figure 5: Scenario configuration of Fuzzy Split-MNIST