Learning To Teach:A Study Of Machine Teaching And Learning In Deep Learning

In recent years,thanks to the contiguous progress of deep learning technology and hardware performance,artificial intelligence develops rapidly and has been widely used in many research areas.Deep learning typically bases on deep artificial neural network and optimized with back propagation algorithm to automatically learn hidden features and patterns from data.Due to the development of deep learning technology,complexity of model architectures and learning algorithms has gradually increased,and its more complex for researchers to design deep learning process for target task.For a specific deep learning task,the whole process includes data preprocessing,design of model architecture and algorithm,and hyper-parameter fine-tuning.These steps requires both knowledge in deep learning and domain-specific knowledge,resulting in a high threshold for the application of deep learning technology to non-experts.Researchers have proposed several automated machine learning algorithms to automatically learn the learning process for specific tasks,thereby reducing the application threshold of deep learning technology and labor cost of algorithm design.However,traditional automated machine learning algorithms have four key problems:previous researches usually 1)have poor efficiency,which hurts the wide application of algorithms;2)limited to a subarea of deep learning,leading to poor transferability between different areas;3)focus on a small aspect of the deep learning process,which hurts the generality of algorithm;4)only focus on target task,and pay little attention to the part that guides model training(meta model).Due to these problems,this paper conducts a deep study on the area of automated machine learning,and proposes several works to solve these problems.1.For the lack of attention to the meta model,we propose learning to teach(L2T)framework based on the concept of teaching and learning in human society.This framework explicitly introduces the concept of "teacher" and "student",enables communications between teacher and student,and jointly optimizes them.2.For the problem that traditional algorithms are only designed for a small aspect of the deep learning process,we apply the L2T framework on multiple learning steps like data selection,model architecture design and loss function design.Firstly,we propose a data selection algorithm called L2T-Data based on the learning to teach framework.The L2T-Data algorithm connects teachers and students through simple state functions and action vectors,and uses deep reinforcement learning algorithms to automatically learn data selection strategies for specific tasks,and achieves good results in image classification and text classification.Secondly,we propose a model architecture learning algorithm called L2T-Model that optimizes with gradient-based method.During the training process of the teacher model in L2T-Model algorithm,different student architectures are selected considering different training states of the student model and data features to achieve better results that human-design model architectures.3.For the lack of transferability in traditional algorithms,we design the L2T framework to improve the transferability of teacher models between tasks from different deep learning areas,and proved by experiments.4.For the problem of efficiency,we propose a learning to teach optimization algorithm L2T-DI based on reverse mode differentiation.L2T-DI algorithm enables deep interaction between the teacher and the student by including internal information of the student model into the state,and directly optimizes the teacher model by iterative reverse mode differentiation steps.The L2T-DI algorithm improves the convergence speed and training effect of the L2T-based data selection algorithm,and achieves better performance on multiple tasks.Experimental results show that the L2T framework proposed in this paper can achieve better performances than classical human-designed networks and traditional automated machine learning algorithms in tasks such as image classification,text classification and neural machine translation,with acceptable overhead and good transferability between tasks,which prove the effectiveness of L2T.In addition,this paper also applies the L2T framework to real world machine learning scenarios,searches for the best model architecture and hyperparameter setting for machine translation task,and achieves better performance than other neural network machine translation models.