Incremental Learning of Object Detectors without Catastrophic Forgetting

Shmelkov, Konstantin; Schmid, Cordelia; Alahari, Karteek

doi:10.48550/arxiv.1708.06977

Cited by 4 publications

(6 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Different from the standard multi-task learning methods which must jointly learn from the previously observed task data in the offline regime, lifelong learning methods explore how to establish the relationships between the observed tasks with new coming ones, and avoid losing performance among the previously encountered tasks. Generally speaking, by compressing the previous knowledge into a compact knowledge library [2], [17], [38] or storing the knowledge in the learned network weights [22], [36], [51], the major procedure in most existing state-of-the-arts is to transfer knowledge from the current knowledge library [2], [38] to learn a new coming task and congest the fresh knowledge over time; or simple-retrain the deep lifelong learning network via overcoming catastrophic forgetting [22], [34], [39], [51]. Despite the success of lifelong machine learning, the basic assumption for most existing models is that all learned tasks are drawn i.i.d.…”

Section: Introductionmentioning

confidence: 99%

Representative Task Self-Selection for Flexible Clustered Lifelong Learning

Sun

Cong

Wang

et al. 2022

IEEE Trans. Neural Netw. Learning Syst.

View full text Add to dashboard Cite

Consider the lifelong machine learning paradigm whose objective is to learn a sequence of tasks depending on previous experiences, e.g., knowledge library or deep network weights. However, the knowledge libraries or deep networks for most recent lifelong learning models are with prescribed size, and can degenerate the performance for both learned tasks and coming ones when facing with a new task environment (cluster). To address this challenge, we propose a novel incremental clustered lifelong learning framework with two knowledge libraries: feature learning library and model knowledge library, called Flexible Clustered Lifelong Learning (FCL 3 ). Specifically, the feature learning library modeled by an autoencoder architecture maintains a set of representation common across all the observed tasks, and the model knowledge library can be self-selected by identifying and adding new representative models (clusters).When a new task arrives, our proposed FCL 3 model firstly transfers knowledge from these libraries to encode the new task, i.e., effectively and selectively soft-assigning this new task to multiple representative models over feature learning library. Then, 1) the new task with a higher outlier probability will be judged as a new representative, and used to redefine both feature learning library and representative models over time; or 2) the new task with lower outlier probability will only refine the feature learning library. For model optimization, we cast this lifelong learning problem as an alternating direction minimization problem as a new task comes. Finally, we evaluate the proposed framework by analyzing several multi-task datasets, and the experimental results demonstrate that our FCL 3 model can achieve better performance than most lifelong learning frameworks, even batch clustered multi-task learning models.

show abstract

Section: Introductionmentioning

confidence: 99%

Representative Task Self-Selection for Flexible Clustered Lifelong Learning

Sun

Cong

Wang

et al. 2022

IEEE Trans. Neural Netw. Learning Syst.

View full text Add to dashboard Cite

show abstract

“…The first loss term L f ocal and the second loss term L r eдr are the standard loss functions used in [24] to train RetinaNet for the new classes where Y n represents the ground-truth one-hot classification labels, Ŷn represents the new model's classification output over n new classes, B n represents the ground-truth bounding box coordinates, and Bn represents the predicted bounding box coordinates for the ground-truth objects. 1 The third loss term L dist _cl as is the distillation loss for the classification subnet similar to that defined in [22,34]. Here, Y o is the output of the frozen old model F' for m old classes using the new training data, and Ŷo is the output of the new model F for the old classes.…”

Section: Incremental Learning Methodsmentioning

confidence: 99%

“…One important approach is adding regularization [2,19] to the neural network weights according to their importance to the old classes, i.e., it discourages the change on important weights using a smaller learning rate. The other major research direction is based the knowledge distillation [12], which uses the new data to distill the knowledge (i.e., the network output) from the old model and mimic its behavior (i.e., generate similar output) when training the new model [22,31,34]. Among the related work, only [34] has focused on the object detection problem.…”

Section: Related Workmentioning

confidence: 99%

“…The other major research direction is based the knowledge distillation [12], which uses the new data to distill the knowledge (i.e., the network output) from the old model and mimic its behavior (i.e., generate similar output) when training the new model [22,31,34]. Among the related work, only [34] has focused on the object detection problem. However, their approach works on the Fast-RCNN model [8], where bounding boxes proposals are computed using an additional selective search method [36] -this approach will not be able to satisfy the real-time inference requirement on edge devices.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Rilod

Tasci

Ghosh

et al. 2019

Proceedings of the 4th ACM/IEEE Symposium on Edge Computing

View full text Add to dashboard Cite

Object detection models shipped with camera-equipped edge devices cannot cover the objects of interest for every user. Therefore, the incremental learning capability is a critical feature for a robust and personalized object detection system that many applications would rely on. In this paper, we present an efficient yet practical system, RILOD, to incrementally train an existing object detection model such that it can detect new object classes without losing its capability to detect old classes. The key component of RILOD is a novel incremental learning algorithm that trains end-to-end for one-stage deep object detection models only using training data of new object classes. Specifically to avoid catastrophic forgetting, the algorithm distills three types of knowledge from the old model to mimic the old model's behavior on object classification, bounding box regression and feature extraction. In addition, since the training data for the new classes may not be available, a real-time dataset construction pipeline is designed to collect training images on-the-fly and automatically label the images with both category and bounding box annotations. We have implemented RILOD under both edge-cloud and edge-only setups. Experiment results show that the proposed system can learn to detect a new object class in just a few minutes, including both dataset construction and model training. In comparison, traditional * Work was done when the author was a full-time employee with SRA. † Work was done when the author was a summer intern with SRA.fine-tuning based method may take a few hours for training, and in most cases would also need a tedious and costly manual dataset labeling step.

show abstract

“…The inner latent variable inference component of LILAC possesses strong similarities to the continual and lifelong learning setting [18]. Many continual and lifelong learning aim to learn a variety of tasks without forgetting previous tasks [30,58,35,3,37,45,49,40,48]. We consider a setting where it is practical to store past experiences in a replay buffer [41,16].…”

Section: Related Workmentioning

confidence: 99%

Deep Reinforcement Learning amidst Lifelong Non-Stationarity

Xie¹,

Harrison²,

Finn³

2020

Preprint

View full text Add to dashboard Cite

As humans, our goals and our environment are persistently changing throughout our lifetime based on our experiences, actions, and internal and external drives. In contrast, typical reinforcement learning problem set-ups consider decision processes that are stationary across episodes. Can we develop reinforcement learning algorithms that can cope with the persistent change in the former, more realistic problem settings? While on-policy algorithms such as policy gradients in principle can be extended to non-stationary settings, the same cannot be said for more efficient off-policy algorithms that replay past experiences when learning. In this work, we formalize this problem setting, and draw upon ideas from the online learning and probabilistic inference literature to derive an off-policy RL algorithm that can reason about and tackle such lifelong non-stationarity. Our method leverages latent variable models to learn a representation of the environment from current and past experiences, and performs off-policy RL with this representation. We further introduce several simulation environments that exhibit lifelong non-stationarity, and empirically find that our approach substantially outperforms approaches that do not reason about environment shift. 1

show abstract

Incremental Learning of Object Detectors without Catastrophic Forgetting

Cited by 4 publications

References 0 publications

Representative Task Self-Selection for Flexible Clustered Lifelong Learning

Representative Task Self-Selection for Flexible Clustered Lifelong Learning

Rilod

Deep Reinforcement Learning amidst Lifelong Non-Stationarity

Contact Info

Product

Resources

About