Algorithm Research On Knowledge Reuse And Generalization Ability Of Meta-learning

In recent years,as a representative of machine learning methods,deep learning has made important advances in algorithms,theories,and applications and has greatly boosted the development of artificial intelligence technologies.Deep learning technology has demonstrated superior performance by relying on massive,high-quality,labeled data and large-scale computing resources.However,it still performs poorly on tasks that are small in quantity,poor in quality,and lack labeled data.Humans can learn the commonality of methodology by learning from different tasks and quickly learn new skills or adapt to a new environment with only a small amount of information.This extraordinary knowledge reusability(learning to learn)is an important part of human intelligence.Therefore,metalearning,which aims to solve the problem of learning to learn,is considered to be one of the keys to artificial general intelligence.The main goal of meta-learning is to quickly adapt to new tasks with a few samples and to realize end-to-end automatic adjustment for hyperparameters and designing for model structures(highly automated machine learning)when applying machine learning to real problems.Traditional machine learning methods realize the generalization to the dataset of a single task by extracting the inherent laws of a large amount of data.In contrast,meta-learning realizes the generalization to a kind of task by extracting the commonality of methodology in different tasks.As a new research hotspot in recent years,meta-learning has its unique methodological advantages,has made many breakthroughs,and has broad theoretical research value and application prospects.This thesis takes advantage of the knowledge reusability and generalization ability of meta-learning and focuses on the two main challenges of meta-overfitting and task designing.The main research contents include the following aspects:(1)In this thesis,considering the local minimum(saddle point)meta-overfitting problem in the meta-reinforcement learning scenario,we proposed Exploration with Structured Noise in Parameter Space(ESNPS).ESNPS combines the idea of escaping the local minimum(saddle point)by utilizing the noise in a neural network and the way of human exploring by utilizing the prior knowledge.ESNPS effectively uses previous search experience as structured noise,perturbs the model through the parameter space,and shows impressive exploring behaviors guided by prior knowledge in new tasks.In particular,as far as we know,this is the first study that uses the learned meta-parameters of the policy network as structured noise in the parameter space.Experimental results on four groups of tasks: cheetah velocity,cheetah direction,ant velocity,and ant direction demonstrate the superiority of ESNPS against a number of competitive baselines.(2)From the perspective of the diverse requirements of meta-learning tasks,this thesis proposed an unsupervised meta-learning method,Constructing Unsupervised Metalearning tasks with Clustering and Augmentation(CUMCA),that constructs tasks from the unlabeled dataset.CUMCA satisfies the class distinction and class consistent requirements by combining unsupervised embedding learning and data augmentation methods.However,the augmented data will introduce bias and weak diversity issues.Therefore,CUMCA exchanges the data of the inner and outer loops to handle such issues.Subsequently,this thesis provides theoretical analysis to explain why the outer loop of MAML is more sensitive to augmented data than the inner loop.Additionally,a detailed analysis of the effects of using these two types of data in different training modes and how many shots of generated augmented data should be used to obtain better performance is given in the experiments.Furthermore,in order to deal with the problem of bias and insufficient diversity of the augmented data,this thesis proposes a new data augmentation method: Prior-Mixup,which uses extra samples as prior to perturbing the original samples and does not change the label.On the public benchmark dataset,CUMCA defeats other unsupervised meta-learning methods and unsupervised embedding learning methods.(3)Considering that all the existing meta-learning recommendation systems ignore the key meta-memorization overfitting problem,which causes poor generalization ability problem,which leads to poor performance,this thesis proposes a cross-domain metaaugmentation method for content-aware recommendation(Meta CAR).Meta CAR uses data augmentation methods to construct mutually exclusive tasks.Mutually exclusive tasks refer to tasks with the same data but different labels and are the key to deal with meta-memorization overfitting.Meta CAR learns a prior by a cross-domain component and utilizes the already observed user-item pairs from the target domain to generate augmented data that is meaningful but distinguishable from real data.The experiments verify that Meta CAR is superior over many different types of excellent baselines.Compared with other meta-learning recommendation systems,Meta CAR has made substantial progress.