Online Continual Learning on Sequences

Parisi, German I.; Lomonaco, Vincenzo

doi:10.1007/978-3-030-43883-8_8

Cited by 19 publications

(12 citation statements)

References 66 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They also recommend some desiderata and guidelines for future CL research. [27] emphasizes the importance of online CL and discusses recent advances in this setting. Although these three surveys descriptively review the recent development of CL and provide practical guidelines, they do not perform any empirical comparison between methods.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Online Continual Learning in Image Classification: An Empirical Survey

Mai

Jeong

et al. 2021

Preprint

View full text Add to dashboard Cite

Online continual learning for image classification studies the problem of learning to classify images from an online stream of data and tasks, where tasks may include different classes (class incremental) or data nonstationarity (domain incremental). One of the key challenges of continual learning is to avoid catastrophic forgetting (CF), i.e., forgetting old tasks in the presence of more recent tasks. Over the past few years, a large range of methods and tricks have been introduced to address the continual learning problem, but many have not been fairly and systematically compared under a variety of realistic and practical settings.To better understand the relative advantages of various approaches and the settings where they work best, this survey aims to (1) compare stateof-the-art methods such as Maximally Interfered Retrieval (MIR), iCARL, and GDumb (a very strong baseline) and determine which works best at different memory and data settings as well as better understand the key source of CF; (2) determine if the best online class incremental methods are also competitive in domain incremental setting; and (3) evaluate the performance of 7 simple but effective trick such as "review" trick and nearest class mean (NCM) classifier to assess their relative impact. Regarding (1), we

show abstract

Section: Related Workmentioning

confidence: 99%

“…We consider the supervised image classification problem with an online (potentially infinite) non-i.i.d stream of data, following the recent CL literature [17,18,27,14]. Formally, we define a data stream of unknown distributions D = {D 1 , .…”

Section: Problem Definition and Evaluation Settingsmentioning

confidence: 99%

Online Continual Learning in Image Classification: An Empirical Survey

Mai

Jeong

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…We define and evaluate the proactive training for continuously training DL models. In a different setting than ours, this problem is similar to continual learning [21][22][23][24]. In that framework, a model is learning tasks one after another with the goal to learn the last task without catastrophic forgetting [25] that is losing the ability to perform on previous tasks.…”

Section: Related Workmentioning

confidence: 99%

Continuous Training and Deployment of Deep Learning Models

et al. 2021

View full text Add to dashboard Cite

Deep Learning (DL) has consistently surpassed other Machine Learning methods and achieved state-of-the-art performance in multiple cases. Several modern applications like financial and recommender systems require models that are constantly updated with fresh data. The prominent approach for keeping a DL model fresh is to trigger full retraining from scratch when enough new data are available. However, retraining large and complex DL models is time-consuming and compute-intensive. This makes full retraining costly, wasteful, and slow. In this paper, we present an approach to continuously train and deploy DL models. First, we enable continuous training through proactive training that combines samples of historical data with new streaming data. Second, we enable continuous deployment through gradient sparsification that allows us to send a small percentage of the model updates per training iteration. Our experimental results with LeNet5 on MNIST and modern DL models on CIFAR-10 show that proactive training keeps models fresh with comparable—if not superior—performance to full retraining at a fraction of the time. Combined with gradient sparsification, sparse proactive training enables very fast updates of a deployed model with arbitrarily large sparsity, reducing communication per iteration up to four orders of magnitude, with minimal—if any—losses in model quality. Sparse training, however, comes at a price; it incurs overhead on the training that depends on the size of the model and increases the training time by factors ranging from 1.25 to 3 in our experiments. Arguably, a small price to pay for successfully enabling the continuous training and deployment of large DL models.

show abstract

“…In this section, we discuss the recent state-of-the-art OCL strategies shown in Fig. 2, that are developed by a combination of three basic CL strategies, i.e., regularization, architectural, and rehearsal [8]. We also summarize the key characteristics of these methods in Table . 1.…”

Section: Online Continual Learning Strategiesmentioning

confidence: 99%

“…A major challenge in realizing a real-world autonomous system capable of continuously learning and adapting over time is to prevent catastrophic forgetting. The learning model needs to maintain a balance between plasticity (the ability to adapt to new knowledge) and stability (the ability to retain prior knowledge) [7], [8]. Excessive plasticity can cause forgetting the prior learned information while learning a new task.…”

Section: Introductionmentioning

confidence: 99%

Continual Learning for Real-World Autonomous Systems: Algorithms, Challenges and Frameworks

Shaheen¹,

Hanif²,

Hasan³

et al. 2021

Preprint

View full text Add to dashboard Cite

Continual learning is essential for all real-world applications, as frozen pre-trained models cannot effectively deal with non-stationary data distributions. The purpose of this study is to review the state-of-the-art methods that allow continuous learning of computational models over time. We primarily focus on the learning algorithms that perform continuous learning in an online fashion from considerably large (or infinite) sequential data and require substantially low computational and memory resources. We critically analyze the key challenges associated with continual learning for autonomous real-world systems and compare current methods in terms of computations, memory, and network/model complexity. We also briefly describe the implementations of continuous learning algorithms under three main autonomous systems, i.e., self-driving vehicles, unmanned aerial vehicles, and robotics. The learning methods of these autonomous systems and their strengths and limitations are extensively explored in this article.

show abstract

Online Continual Learning on Sequences

Cited by 19 publications

References 66 publications

Online Continual Learning in Image Classification: An Empirical Survey

Online Continual Learning in Image Classification: An Empirical Survey

Continuous Training and Deployment of Deep Learning Models

Continual Learning for Real-World Autonomous Systems: Algorithms, Challenges and Frameworks

Contact Info

Product

Resources

About