Learning After Model Deployment

TL;DR

This paper defines Autonomous Learning after Model Deployment (ALMD), a post-deployment paradigm where an AI agent continually detects and learns new classes with labels supplied by humans, avoiding retraining from scratch. It introduces PLDA, a Post-deployment Learning based on Linear Discriminant Analysis, which freezes a pre-trained feature extractor and a shared covariance $\Sigma$ while incrementally updating class means $\mu_i$ for newly learned classes, enabling continual OOD detection via Mahalanobis-based distances. The method jointly addresses continual OOD detection and class-incremental learning in an online setting, using MD or RMD for OOD scoring and minimizing catastrophic forgetting by avoiding parameter updates. Experimental results on CIFAR-10, CIFAR-100, and TinyImageNet show PLDA achieving accuracy close to joint-training upper bounds with significantly faster post-deployment performance, and outperforming several online CL baselines, even under data-scarce labeling conditions. The work highlights ALMD’s practical potential for open-world AI agents, while acknowledging that further advances in continual OOD detection are needed to reach SOTA in all settings.

Abstract

In classic supervised learning, once a model is deployed in an application, it is fixed. No updates will be made to it during the application. This is inappropriate for many dynamic and open environments, where unexpected samples from unseen classes may appear. In such an environment, the model should be able to detect these novel samples from unseen classes and learn them after they are labeled. We call this paradigm Autonomous Learning after Model Deployment (ALMD). The learning here is continuous and involves no human engineers. Labeling in this scenario is performed by human co-workers or other knowledgeable agents, which is similar to what humans do when they encounter an unfamiliar object and ask another person for its name. In ALMD, the detection of novel samples is dynamic and differs from traditional out-of-distribution (OOD) detection in that the set of in-distribution (ID) classes expands as new classes are learned during application, whereas ID classes is fixed in traditional OOD detection. Learning is also different from classic supervised learning because in ALMD, we learn the encountered new classes immediately and incrementally. It is difficult to retrain the model from scratch using all the past data from the ID classes and the novel samples from newly discovered classes, as this would be resource- and time-consuming. Apart from these two challenges, ALMD faces the data scarcity issue because instances of new classes often appear sporadically in real-life applications. To address these issues, we propose a novel method, PLDA, which performs dynamic OOD detection and incremental learning of new classes on the fly. Empirical evaluations will demonstrate the effectiveness of PLDA.

Figures (3)

• Figure 1: Pipeline of the proposed PLDA method. ID (in-distribution) classes, which are used in OOD detection, include both $C$ and $C^L$.
• Figure 2: Accuracies of PLDA for different $C^L$ threshold values for different datasets. While there is no CF, the ID accuracy drops as expected as more classes are learned, which makes the classification harder. The decrease in the ID accuracy is minimal relative to the significant gains in the OOD accuracy, demonstrating the model’s resilience.
• Figure 3: Run time comparison. PLDA achieves the best performance and the fastest run time among all methods across both setups.