UNSUPERVISED CLASS-INCREMENTAL LEARNING THROUGH CONFUSION

Abstract

While many works on Continual Learning have shown promising results for mitigating catastrophic forgetting, they have relied on supervised training. To successfully learn in a label-agnostic incremental setting, a model must distinguish between learned and novel classes to properly include samples for training. We introduce a novelty detection method that leverages network confusion caused by training incoming data as a new class. We found that incorporating a class-imbalance during this detection method substantially enhances performance. The effectiveness of our approach is demonstrated across a set of image classification benchmarks: MNIST, SVHN, CIFAR-10, CIFAR-100, and CRIB.

1. INTRODUCTION

The development of continually learning systems remains to be a major obstacle in the field of artificial intelligence. The primary challenge is to mitigate catastrophic forgetting: learning new tasks while maintaining the ability to perform old ones. This domain of research is often referred to as Continual Learning, Lifelong Learning, Sequential Learning, or Incremental Learning: each with subtleties in the learning environment and training process, but most with the use of supervision (De Lange et al. (2020) ). Recently, Stojanov et al. (2019) introduced a novel unsupervised class-incremental learning problem motivated by the desire to simulate how children's play behaviors support their ability to learn object models (IOLfCV). Here, sequential tasks take the form of exposures. Each exposure is comprised of a set of images that pertains to a single class that is hidden from the learner. Exposure boundaries, the transition from one exposure to the next, are known. The model trained in this setting, is analogous to a young child that has been placed in a playpen with a set of new toys. The child steadily gathers information over time by picking up, examining, and putting down new/old objects continuously. Similar to how the child does not have any guidance to the toys they will examine, the agent does not have access to the exposure's identity during training. To learn in the unsupervised class-incremental setting, an agent must conduct two procedures successfully. Given a new learning exposure, the key step is to perform novelty detection: to identify whether an exposure corresponds to a class that has been learned. If the agent determines that an exposure is familiar, the second step is to identify its label such that the exposure can be leveraged to update the model. Both procedures must be performed reliably. Otherwise, the novelty detection mistakes will result in label noise that distorts the learned model, increasing the likelihood of subsequent mistakes. 2019) applies a similar method to the unsupervised incremental setting by computing class features produced from a set of supervised samples. In contrast, we propose a model, Incremental Learning by Accuracy Performance (iLAP), that determines class novelty and identity by considering performance changes of previously learned tasks when an incoming set of exposure images are trained under a new label. Instead of using a distance-based metric, our novelty detection threshold relies on the percentage of accuracy that was maintained by performing a model update using the incoming exposure. This poses several practical advantages: First, the threshold value does not rely on supervised samples and is more intuitive (Section 3.3). Second, the performance of our method is independent of the sequence of the incoming exposure classes (Section 5.2). Finally, the model is able to distinguish between similar classes more reliably (Section 5.3). From our experiments, we demonstrate that the confusion resulting from training with label ambiguity results in a more reliable signal for novelty detection in comparison to previous methods. We demonstrate that our technique is more robust and results in substantial performance gains in comparison to various baselines. Furthermore, despite the absence of labels, our model was able to perform similarly to supervised models under several benchmarks. In summary, this work provides three contributions: • We present a novel framework, iLAP, that achieves learning in the unsupervised classincremental environment where the exposure identities are unknown. • We demonstrate that by including a class-imbalance technique, our unsupervised method is able to closely match supervised performance for several image classification benchmarks trained in the incremental setting. • We identify failure cases that are overlooked by traditional OOD methods that leverage distance-based thresholds.

2. RELATED WORKS

Introduced by Stojanov et al. ( 2019), the unsupervised class-incremental setting contains a set of sequential tasks that are single-class exposures; classes pertaining to the exposures may repeat and are unknown. This is not to be mistaken with unsupervised continual learning (UCL) where task boundaries and task identities are unavailable (Lee et al. 2017)). In contrast to prior works, our method addresses incremental learning in a setting where exposure labels are unavailable.

2.2. UNSUPERVISED CONTINUAL LEARNING

Recently, a series of works tackle the UCL problem where task boundaries and task identities are unknown. Smith & Dovrolis (2019) performs novelty detection by analyzing an input image through a series of receptive fields to determine if an input patch is an outlier. Meanwhile, CURL proposes a method to learn class-discriminative representations through a set of shared parameters (Rao et al. ( 2019)). CN-DPM, introduces an expansion-based approach that utilizes a mixture of experts to learn feature representations (Lee et al. ( 2020)). Although CN-DPM performs in a task-free setting, incoming tasks are multi-class and individual class labels are provided. This supervision is required to train the existing experts and determine when a new one is needed. While boundary information is not required for these works, the performances are far below supervised baselines (77.7% on and MNIST 13.2% Omniglot) (Rao et al. (2019) ).



Deep neural networks are known to make overconfident decisions for anomalous data distributions that were not seen during training (Hendrycks & Gimpel (2016)). To address this problem, research related to out-of-distribution (OOD) detection have utilized supervised methods (Liang et al. (2017); Alemi et al. (2018)) and unsupervised methods (Choi & Jang (2018); Hendrycks et al. (2018); Serrà et al. (2019)). Works related to open set recognition have also addressed the OOD problem by applying distance-based thresholds computed from known class scores (Scheirer et al. (2012; 2014)). The work by Stojanov et al. (

(2020); Smith & Dovrolis (2019); Rao et al. (2019)). Our work presents an agent that is able to leverage the boundary information from the unsupervised class-incremental environment to achieve performances that are close to models trained under supervision. 2.1 CONTINUAL LEARNING/INCREMENTAL LEARNING Prior works in this field primarily aim to improve a model's ability to retain information while incorporating new tasks (Goodfellow et al. (2013); Parisi et al. (2019); Rebuffi et al. (2017); Lopez-Paz & Ranzato (2017); Aljundi et al. (2018); Castro et al. (2018)). Typically, these models reside in learning settings where both task labels and task boundaries are available. Methods include replay techniques, the usage of artifacts and generated samples to refresh a model's memory (Kamra et al. (2017); Wu et al. (2018); Rolnick et al. (2019); Shin et al. (2017); Wu et al. (2019)), and regularization-based practices, the identification and preservation of weights that are crucial for the performance of specific tasks (Kirkpatrick et al. (2017); Zenke et al. (2017); Yoon et al. (

