Representation Learning And Incremental Technology For Data-Efficient Machine Learning

Modern machine learning methods have achieved promising performance in various public datasets or real-world(open-world)applications.Currently,in the standard paradigm of machine learning,data is the cornerstone of this field.However,a point that needs attention is that the powerful performance presented by the current learning paradigm relies on a large amount of data as support,and this data hunger behavior is also included as an unresolved issue in general AI.Second,data privacy has also been a long-standing and widespread concern in the machine learning community.Therefore,how to efficiently use data for machine learning has become a priority research direction.Data-efficient machine learning methods refer to the use of limited data to train machine learning models,which perform the learning process in limited data scenarios while still achieving high accuracy and generalization capabilities.This paradigm focuses on the positive interaction between algorithms and data,which reduce the demand and overhead of data for machine learning methods,thus making machine learning more secure,convenient,and efficient.In this thesis,we focus on two key technologies in data-efficient machine learning,namely representation learning and incremental learning.We aim to efficiently utilize data resources to improve the performance and generalization ability of models,thereby better addressing visual tasks in complex scenarios.First,the current approaches for person identification utilizing the entire image deep features only are not sufficient to preserve identities due to the reliance on visible visual representations.In this thesis,we propose a novel person representation method using a graph-powered pose representation to encode image representations at the level of relational induction bias.Our framework consists of two modules:(i)a novel pose-guided representation module that is able to capture the pose changes dynamically and their associated effects;(ii)a pose-guided graph embedding module using both the image deep features and the pose structure information for a better person representation inference.We use the obtained representations as an identity cue and apply it to solve the real-world problem of player identification in broadcast videos.Experiment results on the real-world sport game scenarios demonstrate that our method achieves state-of-the-art identification performance,together with a better player representation.The proposed method mines the structural information in the data,achieves better performance and realizes a data-efficient representation technique.Second,in Incremental Object Detection(IOD),which relies on currently available data,previous work mainly focuses on distilling for the combination of features and responses.However,they under-explore the information that contains in responses.In this thesis,we propose a response-based incremental distillation method,dubbed Elastic Response Distillation(ERD),which focuses on elastically learning responses from the classification head and the regression head.Firstly,our method transfers category knowledge while equipping student detector with the ability to retain localization information during incremental learning.In addition,we further evaluate the quality of all locations and provide valuable responses by the Elastic Response Selection(ERS)strategy.Extensive experiments conducted on MS COCO demonstrate our method achieves state-of-the-art result,which substantially narrows the performance gap towards full training.The proposed method exploits the response information in the detection task under the limited data authority,and realizes a data-efficient incremental detection technology.Third,learning from changing tasks and sequential experience without forgetting the obtained knowledge is a challenging problem for artificial neural networks.In this work,we focus on two challenging problems in the paradigm of Continual Learning(CL)without involving any old data:(i)the accumulation of catastrophic forgetting caused by the gradually fading knowledge space from which the model learns the previous knowledge;(ii)the uncontrolled tug-ofwar dynamics to balance the stability and plasticity during the learning of new tasks.In order to tackle these problems,we present Progressive Learning without Forgetting(PLwF)and a credit assignment regime in the optimizer.PLwF densely introduces model functions from previous tasks to construct a knowledge space such that it contains the most reliable knowledge on each task and the distribution in-formation of different tasks,while credit assignment controls the tug-of-war dynamics by removing gradient conflict through projection.Extensive ablative experiments demonstrate the effectiveness of PLwF and credit assignment.The proposed method explores the distribution information of the data under the limited data authority,and realizes a data-efficient incremental learning technology.This thesis focuses on the positive interaction between algorithms and data from the perspective of representation learning and incremental learning,aiming to achieve data-efficient machine learning.With the limited data authority,the proposed method efficiently mines data information,improving the sample-based learning ability and the continuous learning level of models in the standard machine learning paradigm.