Research On Continual Learning Of Convolutional Neural Networks Inspired By Human Brain Learning And Memory Mechanisms

In recent years,convolutional neural networks have achieved great success in the field of artificial intelligence.In individual tasks such as image classification and Go games,they have reached the level of human beings,and even surpassed human beings.But the convolutional neural network at this stage is rooted in the closed world assumption,that is,the model is learned on a closed,static,and identically distributed training set,and then reasoned in the unknown open world.Obviously,the network based on this learning paradigm cannot continuously learn new knowledge from the task flow,and it is difficult to adapt or expand its behavior over time,so this kind of intelligence is very limited.Human beings can learn from experience and continuously reuse and expand the learned knowledge to new environments.This ability is called continual learning or lifelong learning.Continual learning is an important feature of an agent with strong artificial intelligence or general artificial intelligence,and this thesis reviews the continual learning from the perspective of human-intelligence level.Catastrophic forgetting is a basic issue in continual learning of convolutional neural networks.Catastrophic forgetting refers to a phenomenon in which the performance of old knowledge drops sharply after the agent learns new knowledge.Catastrophic forgetting causes convolutional neural networks to only master a few tasks at the moment or in the immediate vicinity,hindering it to accumulate knowledge incrementally.Under the continuous learning paradigm,a steady stream of tasks are sequentially input into the convolutional neural network for training,and the network dynamically updates the model parameters under the guidance of the optimization strategy according to the data representation of each task.There are three problems that cause or exacerbate catastrophic forgetting:(1)The accuracy of parameter importance measurement in the model:the parameters in the model contain information about the task,and keeping important parameters in the learning process can consolidate knowledge,Therefore,whether the parameter importance can be measured accurately is important to alleviate catastrophic forgetting.(2)The problem of data similarity in continuous tasks:the excessive similarity of data in the task will lead to blurred decision boundaries between the two,making the network unable to learn the key features of the category.Therefore,whether the similarity between tasks can be reduced is important to alleviate catastrophic forgetting.(3)The stability of the local minimum in optimization:If the objective function achieves a local minimum with low stability in the optimization space,the network is more likely to deviate from its optimal parameter space in subsequent updates.Therefore,whether the local optimum is smooth is important to alleviate catastrophic forgetting.Inspired by the human memory mechanism,this paper designs an algorithm from the three elements involved in the learning paradigm referring to model,data,and training strategy to alleviate the problem of catastrophic forgetting and realize continuous learning of convolutional neural networks.The specific content is as follows:(1)Reviewing continual learning from the perspective of human-level intelligenceThere are currently many reviews on continuous learning,but they generally tend to summarize a specific aspect of continuous learning problems,such as algorithms,applications,or evaluation indicators.However,the purpose of continuous learning of convolutional neural networks is to approach human high-order intelligence,and the above review is insufficient to examine existing research progress from a more fundamental and comprehensive perspective.The complete continuous learning process of humans consists of three parts:information review,information foresight,and information transfer.Among them,information review refers to remembering previously learned information,information foresight refers to continuously learning and inferring new information,and information transfer refers to transferring useful knowledge between different pieces of information.This article is based on the commonalities and connections between artificial neural networks and convolutional neural networks.It summarizes and analyzes existing continuous learning algorithms,evaluation indicators,and applications from three perspectives:information review,information foresight,and information transfer,in order to demonstrate the current achievements in continuous learning of convolutional neural networks and their gap with higher-order intelligent agents such as the human brain.(2)The second-order parameter importance estimation method of convolutional neural network based on neuron death behaviorFrom the perspective of the model,catastrophic forgetting occurs because the parameters of the old task are changed or overwritten by those of the new task,causing the network to lose important features of the old task.Existing parameter-based regularization methods measure parameter importance through the first derivative of the objective function which is also referred to the gradient,but the gradient only reflects the local properties of the objective function,and does not fully reflect all the characteristics of the parameters on the task,so there is a problem of inaccurate estimation of the importance.Inaccurate parameter importance will cause the model to lose key knowledge,which is not conducive to overcoming catastrophic forgetting.The mechanism of synaptic plasticity believes that the accurate balance of human brain memory and loss depends on repeated stimulation and continuous inactivation of cascade neurons.Inspired by this,this paper proposes a second-order important parameter estimation method based on neuron death behavior,which supplements the information of the second-order derivative on the basis of the first-order derivative of the objective function.The basic assumption of the algorithm is that after a parameter disappears,if the loss of the model increases significantly,its importance is greater,and vice versa.Based on this,the algorithm obtains the loss value by preserving the the first and the second derivative of the Taylor expansion on loss function,and change the Hessian matrix into Fisher matrix by Gauss-Newton method to obtain the final parameter importance.The algorithm in this paper has achieved good continuous learning effect in the field of image generation and image rain removal.(3)The similarity data discriminant method of convolutional neural network based on representation separationThe second factor that causes catastrophic forgetting in convolutional neural networks is similar data in continuous tasks.The data in continuous learning has a sequential order,and when the posterior data is similar to the anterior data,the catastrophic phenomenon will be more severe.The occurrence of this phenomenon is because similar data is intertwined in the feature space,which means there is a significant overlap in feature distribution.Therefore,subsequent tasks are easy to cover similar pre order tasks,leading to the model losing key information in the pre order tasks.Existing research has seriously ignored the problem.Interference theory holds that competing information tends to interfere with each other and cause forgetting,and one of the ways to overcome memory interference is to increase the intensity of learning the characteristics of different information.Inspired by this,this paper proposes a similar data discriminant method for convolutional neural networks based on representation separation.This algorithm converts the confusion of data in similar tasks into separability by constructing the same sample set and similar sample set for each batch in the current task,and designs a similar separation function to minimize the feature distance between the same sample set data and maximize the feature distance between similar sample set data,thus forcing the network to learn the most discriminative features in similar categories.The algorithm in this paper achieves a good continuous learning effect in the field of image classification.(4)Self-evolutionary optimization method of convolutional neural network based on hierarchical training from easy to difficultThe third reason for catastrophic forgetting of the convolutional neural network is the training strategy in the continuous learning process.The smaller the stationarity of the local optimal point,the greater the range of changes in the parameters in the model in subsequent updates,and the easier it is to cause representation drift,so the greater the degree of forgetting.Existing studies have seriously ignored this stationarity.Human beings have a behavioral shaping mechanism-the way of learning from easy to difficult,step by step.Inspired by this,this paper proposes an autonomous evolutionary optimization method for convolutional neural networks based on hierarchical training from simple to difficult.This method considers that the shaping learning mechanism from easy to difficult can help the convolutional neural network converge to a local minimum point with greater stability.Based on this,the algorithm in this paper divides the data into shallow data and deep data according to the feature density,and trains the model sequentially through step-by-step learning.Among them,the shallow data provides the basic feature information of the category,and gradually adding the deep data with more complex features helps the model learn the global information.The algorithm in this paper has achieved a good continuous learning effect in the field of image classification and image generation.