WO2022247151A1 - 一种基于大脑机理的认知学习方法 - Google Patents

一种基于大脑机理的认知学习方法 Download PDF

Info

Publication number
WO2022247151A1
WO2022247151A1 PCT/CN2021/129297 CN2021129297W WO2022247151A1 WO 2022247151 A1 WO2022247151 A1 WO 2022247151A1 CN 2021129297 W CN2021129297 W CN 2021129297W WO 2022247151 A1 WO2022247151 A1 WO 2022247151A1
Authority
WO
WIPO (PCT)
Prior art keywords
cognitive
learning
algorithm
case
hyperparameters
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/129297
Other languages
English (en)
French (fr)
Inventor
吴启晖
阮天宸
赵世瑾
周福辉
黄洋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to US17/786,564 priority Critical patent/US11948092B2/en
Publication of WO2022247151A1 publication Critical patent/WO2022247151A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the invention belongs to the field of artificial intelligence, and further relates to a cognitive learning framework and method based on brain mechanism in machine learning.
  • Meta-learning is a learning framework inspired by educational psychology, which has been applied in the field of machine learning.
  • the main components of this framework are meta-features and meta-knowledge base.
  • Meta-learning extracts problem features, runs different machine learning algorithms on a wide range of learning tasks, and collects their performance to form a performance knowledge base.
  • meta-learning has been widely used to solve algorithm selection and hyperparameter optimization problems in classification, image recognition, etc.
  • the disadvantage of this framework is that when faced with a problem that has not appeared in the meta-knowledge base or has a low similarity with existing cases, the performance of meta-learning is often poor.
  • a large prior task set is required, and tasks are assumed to be independent and identically distributed. Once the dynamic change of the task leads to the non-IID characteristics, the framework cannot adapt to this change in time.
  • this framework cannot improve the learning model by exploiting the performance information from real tests, thus, it is difficult to handle complex data and complex learning environments.
  • the present invention aims at the deficiencies of the above-mentioned prior art, and proposes a cognitive learning method based on brain mechanism.
  • the cognitive feature extraction module extracts the features of the dynamic environment and dynamic tasks, and transmits them to the cognitive control module;
  • the cognitive control module establishes a mapping relationship between the features obtained in step (1b) and suitable algorithms and hyperparameters, so as to obtain suitable algorithm types and hyperparameter combinations;
  • the cognitive control module calls the specific algorithm type and hyperparameter combination selected in step (1c) from the algorithm and hyperparameter library of the memory module, and transmits it to the learning network;
  • Step (1e) Learning network module executes the algorithm and hyperparameters called in step (1d), executes hyperparameters and algorithms in the dynamic environment and dynamic tasks extracted in step (1b), obtains corresponding learning results, and outputs them to the outside world in a dynamic environment;
  • step (1b) Store the raw data of the dynamic environment and dynamic tasks extracted in the online learning step (1b) to the database in the memory module, and store the characteristics of the dynamic environment and tasks, the algorithm and hyperparameters selected in step (1c) and The learning result obtained in step (1e) is stored as a new cognitive case in the cognitive case library in the memory module;
  • the cognitive control module establishes a mapping relationship between the characteristics of the case in step (2b) and a suitable algorithm and hyperparameters, so as to obtain a suitable combination of algorithm types and hyperparameters;
  • the cognitive control module calls the specific algorithm type and hyperparameter combination selected in step (2c) from the algorithm and hyperparameter library of the memory module, and transmits it to the learning network;
  • the learning network executes the algorithm and hyperparameters obtained in step (2d) based on the dynamic environment and tasks in step (2b), obtains corresponding learning results, and passes them to the cognitive evaluation module;
  • Cognitive case library transmits step (2b) the past best learning result of this case to the cognitive evaluation module
  • the cognitive evaluation module compares the new learning result of the case in step (2e) with the historical best learning result of the case in step (2f), and uses the algorithm and hyperparameters that achieve better learning results and the learning result as The updated historical best algorithm, hyperparameters, and learning performance are passed to the cognitive case library, and the cognitive case in step (2b) is updated;
  • (2h) Retrain the cognitive control module based on the updated cognitive case library.
  • Step (1b) The cognitive feature extraction module extracts the feature f(e, x) of the dynamic environment and dynamic tasks, and sends it to the cognitive control module.
  • Step (1c) uses the cognitive control module to establish a mapping relationship S[f(e, x)] between the features obtained in step (1b) and suitable algorithms and hyperparameters to obtain suitable algorithm types a ⁇ A and hyperparameter combination ⁇ , where A represents the algorithm library, a represents an algorithm, A represents the hyperparameter library, ⁇ represents one of the hyperparameter combinations, and ⁇ is the belonging number, representing the dynamic environment And dynamic tasks come from external databases.
  • the cognitive case library described in step (2a) is used to store the original data [e, x] of the dynamic environment and tasks extracted in the online learning step (1b) to the database in the memory module And take the characteristics f(e, x) of the dynamic environment and tasks, the algorithm and hyperparameters (a, ⁇ ) selected in step (1c) and the learning result y obtained in step (1e) as a new cognitive case [f (e, x), (a best , ⁇ best ), y best ] are stored in the cognitive case library in the memory module.
  • step ( 2c ) The cognitive control module described in step ( 2c ) is used to establish a mapping relationship between the characteristics of the case in step (2b) and the appropriate algorithm and hyperparameters to get the appropriate algorithm type combined with hyperparameters
  • step (2g) compare the new learning outcomes for this case in step (2e) and the historical best learning result for this case in step (2f)
  • the algorithm and hyperparameters with better learning results and the learning results are passed to the cognitive case library as the updated historical best algorithm, hyperparameters and learning performance, and the cognitive case in step (2b) is updated
  • the framework includes four main modules, namely, cognitive feature extraction, cognitive control, learning network and memory module.
  • the memory module contains a database, a cognitive case library, an algorithm and a hyperparameter library, which store data of dynamic environments and tasks, cognitive cases, and specific algorithms and hyperparameter values.
  • the most suitable algorithm type and hyperparameter combination can be flexibly selected.
  • correctly marked data is used to correct wrongly marked data to achieve robustness to training data.
  • the present invention has the following advantages:
  • the present invention introduces the cognitive control module and the memory module by referring to the brain mechanism, the learning results of the cognitive cases stored in the memory module can be called, and the cognitive control module updates the environment and task characteristics and the selected algorithm and hyperparameters In this way, the learning ability can be continuously improved and more suitable algorithms and parameters can be selected. It overcomes the disadvantage that once the existing framework is trained, the learning performance will not be improved, so that the present invention has the ability of self-learning.
  • the present invention introduces the cognitive feature extraction module by referring to the brain mechanism, it can sense the change and extract new features of the dynamic environment and task when the environment or task changes.
  • the cognitive control module changes the selection of algorithm type and hyperparameter combination according to new characteristics according to the matching relationship between environment and task characteristics and algorithm and hyperparameter selection. It overcomes the shortcomings of existing frameworks that cannot adapt to dynamically changing scenarios due to fixed algorithms and hyperparameters, and performs poorly in new scenarios. This enables the present invention to select the most suitable algorithm and hyperparameters for different scenarios, and has the ability to adapt to dynamic environments and tasks.
  • the present invention introduces the cognitive evaluation module by referring to the brain mechanism, it is possible to evaluate the performance of the selected algorithm type and hyperparameters in a certain scene during offline self-learning, and compare the current learning results with the cognitive case Compare the results in , correcting mislabeled training samples. It overcomes the deficiency that the existing framework relies heavily on the quality of training samples, so that the present invention can be robust to mislabeled training samples, reduce bad training samples, and improve learning performance. This makes the present invention have the ability of "good money driving out bad money”.
  • Fig. 1 is a flow chart of the present invention
  • Fig. 2 (a) to Fig. 2 (d) are the comparison charts of the algorithm and hyperparameter selection accuracy rate and image recognition accuracy rate when the number of test sequences is continuously increasing using the present invention and the prior art;
  • Fig. 3 (a) and Fig. 3 (b) are the algorithm and hyperparameter selection accuracy rate comparison chart when the environment and tasks are continuously and dynamically switched using the present invention and the prior art;
  • Fig. 4(a) to Fig. 4(d) are comparison charts of the algorithm, hyperparameter selection accuracy rate and image recognition accuracy rate under different error labeling training sample sizes using the present invention and the prior art.
  • the online learning process corresponds to the executive control process of the brain's cognitive mechanism.
  • Executive control is a complex cognitive process in which individuals dynamically and flexibly regulate the activities of multiple cognitive subsystems during goal-oriented behavior. It operates on external tasks and data in two ways, making plans to guide behavior and quickly adjusting when unexpected changes occur. This mechanism enables primates to rapidly switch behaviors in dynamic environments.
  • e represents the data describing the dynamic environment
  • x represents the data describing the dynamic task
  • d represents the vector composed of these two
  • D represents the external database
  • ⁇ represents the belonging number, indicating that the dynamic environment and dynamic task come from the external database.
  • the cognitive feature extraction module corresponds to the sensory cortex in the cerebral cortex, extracts the feature f(e, x) of the data, and transmits it to the cognitive control module.
  • the cognitive control module corresponds to the prefrontal cortex in the cerebral cortex, integrates the characteristics of the sensory cortex, and performs cognitive control based on the characteristic information and relevant experience knowledge, and obtains the correlation information between the stimulus and the response.
  • A represents the algorithm library
  • a represents one of the algorithms
  • A represents the hyperparameter library
  • ⁇ represents one of the hyperparameter combinations
  • ⁇ is the belonging number, indicating that the algorithm comes from the algorithm library
  • the hyperparameter combination comes from the hyperparameters library.
  • Step 1.4 the memory module corresponds to the memory center in the brain, and the cognitive control module calls the specific algorithm type and hyperparameter combination selected in step 1.3 from the algorithm library A and hyperparameter library ⁇ of the memory module, and passes it to the learning network .
  • the offline self-learning process corresponds to the introspection process of the brain's cognitive mechanism. Since performing the control process is a rapid process, inappropriate motor responses may result. By introspecting events that lead to inappropriate motor responses stored in internal memory, the introspective process can execute more appropriate motor responses.
  • the introspective process is a spontaneous mental activity that is not related to the current task or perceived environment.
  • the memory module corresponds to the memory center in the brain, and stores the original data [e, x] of the dynamic environment and tasks extracted in the online learning step 1.2 to the database in the memory module And take the characteristics f(e, x) of the dynamic environment and tasks, the algorithm and hyperparameters (a, ⁇ ) selected in step 1.3, and the learning result y obtained in step 1.5 as a new cognitive case [f(e, x ), (a best , ⁇ best ), y best ], stored in the cognitive case library in the memory module.
  • the cognitive control module corresponds to the prefrontal cortex in the cerebral cortex, integrates the characteristics of the sensory cortex, and performs cognitive control based on the characteristic information and relevant experience knowledge, and obtains the correlation information between the stimulus and the response. Establish a mapping relationship between the characteristics of the case and the appropriate algorithm and hyperparameters to get the appropriate algorithm type combined with hyperparameters
  • step 2.4 the cognitive control module calls the specific algorithm type and hyperparameter combination selected in step 2.3 from the algorithm and hyperparameter library of the memory module, and transmits it to the learning network.
  • Step 2.5 the learning network module corresponds to the premotor cortex, and a specific motor plan is obtained according to the association information of the stimulus and the response, based on the dynamic environment and the task in step 2.2 Execute the algorithm and hyperparameters obtained in step 2.4 to obtain the corresponding learning results and pass it to the cognitive assessment module.
  • step 2.6 the cognitive case library in the memory module stores the best past learning results of the case in step 2.2 Passed to the Cognitive Assessment Module.
  • the cognitive assessment module corresponds to the anterior cingulate cortex in the cerebral cortex, monitors conflict responses in the premotor cortex, compares the new learning results for this case in step 2.5 and the historical best learning result for this case in step 2.6 Then, the conflict information is fed back to the prefrontal cortex, that is, the algorithm and hyperparameters for better learning results and the learning results are used as the updated historical best algorithm, hyperparameters, and learning performance, and passed to the cognitive case library.
  • the update steps The cognitive case in 2.2 If the historical best learning result y′ * of the sampled case is better than the new learning result y′, it will not be updated. Otherwise, update as follows:
  • Step 2.8 retrain the cognitive control module according to the updated cognitive case library.
  • the simulation experiment of the present invention is carried out under MATLAB R2016a software.
  • An example of the invention is the task of image classification.
  • the example has 5 different environments and 5 data sets, including public data sets MNIST, FashionMNIST, and Cifar-10.
  • ImageNet which takes 5 categories from ImageNet, and each category has 4500 training samples.
  • MNIST with only 1000 training samples per class.
  • mini-ImageNet and mini-MNIST respectively.
  • the cognitive characteristics of the instance include five data set features: the total number of training samples of the images in the data set, the number of image classification categories in the data set, the width of the images in the data set (the images are adjusted to be equal in length and width), the color features of the images in the data set, and the image Classification difficulty; and two performance requirements characteristics: the requirement of image classification accuracy; the requirement of image classification completion time.
  • the cognitive control module of the example adopts the structure of radial basis function neural network (RBFNN) for algorithm selection, and adopts the structure of BP neural network for hyperparameter selection.
  • RBFNN radial basis function neural network
  • the algorithm library of the example contains five different classification algorithms: support vector machine, Adaboosting, random forest, decision tree, and deep neural network.
  • the hyperparameter library contains important hyperparameters for each of these algorithms.
  • Fig. 2 (a) to Fig. 2 (d) are the comparison charts of the algorithm and hyperparameter selection accuracy rate and image recognition accuracy rate when the number of test sequences is continuously increasing using the present invention and the prior art;
  • Fig. 3 (a) and Fig. 3 (b) are the algorithm and hyperparameter selection accuracy rate comparison chart when the environment and tasks are continuously and dynamically switched using the present invention and the prior art;
  • Fig. 4(a) to Fig. 4(d) are comparison charts of the algorithm, hyperparameter selection accuracy rate and image recognition accuracy rate under different error labeling training sample sizes using the present invention and the prior art.
  • Accompanying drawing 2 (a) and accompanying drawing 2 (b) are respectively adopting the present invention and prior art when the number of test sequences increases continuously the algorithm selection accuracy rate and the image classification accuracy rate comparison chart based on this algorithm.
  • Algorithm selection accuracy evaluation is done by checking whether the selected algorithm is the best algorithm for the mark.
  • the evaluation of image classification accuracy is by checking whether the classification results of image classification performed according to the selected algorithm and its optimal hyperparameters are correct.
  • the abscissa indicates the number of test sequences (times), and the ordinate indicates the algorithm selection/image classification accuracy (percentage).
  • the broken line marked with the red square represents the accuracy rate of algorithm selection/image classification using the method of the present invention.
  • the broken line marked with a black circle represents the algorithm selection/image classification accuracy of the state-of-the-art from a library with 30 meta-cases for each dataset
  • the broken line marked with a blue positive triangle represents the state-of-the-art from a library with 30 meta-cases for each dataset.
  • broken line marked with a green inverted triangle represents prior art algorithm selection from library implementation with 250 meta-cases for each dataset Algorithm selection/image classification accuracy.
  • Accompanying drawing 2 (c) and accompanying drawing 2 (d) respectively adopt the present invention and prior art when the number of test orders is constantly increasing the deep neural network hyperparameter selection accuracy rate and the deep neural network image classification accuracy based on this hyperparameter rate comparison chart.
  • Accuracy assessment of hyperparameter selection for deep neural networks is done by checking whether the selected hyperparameters are the best hyperparameters marked.
  • the abscissa indicates the number of test sequences (times), and the ordinate indicates the hyperparameter selection/image classification accuracy (percentage).
  • the broken line marked with a red square represents the hyperparameter selection/image classification accuracy rate of the method of the present invention.
  • the present invention can store the tested cases to the cognitive case space, during the six tests, its accuracy for each data set
  • the number of training samples are 50, 100, 150, 200, 250, 300 respectively.
  • the polyline marked with a black circle represents the state-of-the-art hyperparameter selection/image classification accuracy from a library with 50 meta-cases for each dataset
  • the polyline marked with a blue positive triangle represents the state-of-the-art from Hyperparameter selection/image classification accuracy for hyperparameter selection achieved by a library with 100 meta-cases per dataset
  • the broken line marked with a green inverted triangle represents the state-of-the-art from a library with 250 meta-cases per dataset Hyperparameter selection/image classification accuracy for hyperparameter selection.
  • Accompanying drawing 3 (a) is the comparison chart of the accuracy rate of algorithm selection when the environment and tasks are continuously and dynamically changed using the present invention and the prior art.
  • the abscissa indicates the number of test sequences (times), and the ordinate indicates the algorithm selection accuracy (percentage).
  • the broken line marked with a red square represents the accuracy rate of algorithm selection using the method of the present invention.
  • the broken line marked with a blue triangle represents the state-of-the-art algorithm selection accuracy from a library implementation with 50 time-first meta-cases for mini-ImageNet.
  • the algorithm selection accuracy rate of the method of the present invention and the existing method at the beginning drops to very low, but as the number of tests increases, the method of the present invention The method is rapidly improved, but the accuracy rate of algorithm selection obtained by using the existing technology is always low.
  • Accompanying drawing 3 (b) is the comparison chart of hyperparameter selection accuracy when using the present invention and the prior art when environments and tasks are continuously and dynamically switched.
  • the abscissa indicates the number of test sequences (times), and the ordinate indicates the accuracy rate (percentage) of hyperparameter selection.
  • the broken line marked with a red square represents the accuracy rate of hyperparameter selection using the method of the present invention.
  • the broken line marked with a blue triangle represents the state-of-the-art hyperparameter selection accuracy from a library with 50 time-first meta-cases for mini-ImageNet.
  • the hyperparameter selection accuracy rate of the method of the present invention and the existing method at the beginning are all reduced to very low, but with the increase of the number of tests
  • the inventive method has improved rapidly, but the accuracy rate of hyperparameter selection obtained by using the existing technology has always been low.
  • the broken line marked with a blue positive triangle represents the algorithm selection/image classification accuracy rate of the prior art when the bad coin rate is 10%
  • the broken line marked with a green inverted triangle represents the algorithm selection/image classification rate of the prior art when the bad coin rate is 30%.
  • Image classification accuracy By comparing the changes of algorithm selection/image classification accuracy obtained by the two methods, it can be seen that the algorithm selection/image classification accuracy obtained by the present invention is better than the existing method. Both the present invention and existing methods are initially affected by bad coins, but the performance of the present invention will gradually improve. This is because, on the one hand, better matching relationships between environment and task features and algorithms and hyperparameters are stored in the cognitive case space, and the proportion of mislabeled samples decreases.
  • the cognitive evaluation module can compare the current learning result with the previous learning result, and feed back the evaluation result to the cognitive control module.
  • the existing technology is always affected by wrongly labeled samples, and the accuracy of the algorithm selection/image classification obtained is almost unchanged.
  • the broken line marked with a blue positive triangle represents the hyperparameter selection/image classification accuracy of the state-of-the-art when the bad coin rate is 10%
  • the broken line marked with a green inverted triangle represents the hyperparameters of the state-of-the-art when the bad coin rate is 30% Selection/image classification accuracy.
  • the cognitive evaluation module can compare the current learning result with the previous learning result, and feed back the evaluation result to the cognitive control module.
  • the existing technology is always affected by mislabeled samples, and the accuracy of hyperparameter selection/image classification selection is almost unchanged.
  • the cognitive learning framework and method based on the brain mechanism proposed by the present invention has the ability of self-learning, adapting to dynamically changing environments and tasks, and good money driving out bad money, which makes the present invention practical can be better applied.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供了一种基于大脑机理的认知学习方法，本发明根据环境和任务选择最合适的算法类型和参数，能在各种不同的环境和任务中获得良好的学习效果，并且能修正错误行为。该框架内部包括四个主要的模块，即认知特征提取，认知控制，学习网络以及记忆模块。其中记忆模块内部包含数据库，认知案例库，算法和超参数库，分别存储动态环境和任务的数据，认知案例以及具体的算法和超参数值。针对动态的环境和任务，可以灵活地选择出最合适的算法类型和超参数组合。同时通过"良币驱逐劣币"，用正确标记数据纠正错误标记数据，实现对训练数据的鲁棒性。

Description

一种基于大脑机理的认知学习方法 技术领域
本发明属于人工智能领域，更进一步涉及机器学习中的一种基于大脑机理的认知学习框架与方法。
背景技术
近年来，机器学习受到了极大的关注，成为学术界和工业界研究热点，在众多领域得到广泛应用。人们提出了多种机器学习框架，使得一个系统能够智能地从以往获取的经验中学习。现有的框架有一定优势，然而传统的机器学习框架往往依赖大量的专家知识和大量的高质量标记样本。一旦缺少这些先验，它们的性能将会受到很大的影响。此外，合适的算法模型和参数组合对于学习算法的性能也至关重要。在传统的机器学习算法中，一旦完成了训练过程，模型和超参数将不会再改变，机器学习算法将只能实现训练好的固定的功能，而不能通过自我学习来适应动态的环境和任务。由于这些缺陷，它们的性能和应用范围受到了一定的限制。因此，构建一个能适应动态环境和动态任务的学习框架具有重要的意义和实际应用价值。
P.Kerschke，H.H.Hoos，F.Neumann，H.Trautmann等人在其发表的论文“Automated algorithm selection：Survey and perspectives”(Evolutionary computation，vol.27，no.1，pp.3-45，2019.)中提出了一种统一的元学习算法框架。元学习是一种受教育心理学启发的学习框架，目前已经被应用于机器学习领域。这个框架最主要的组成部分是元特征和元知识库。元学习提取问题特征，在广泛的学习任务上运行不同的机器学习算法，收集它们的性能构成性能知识库。通过从这种经验(也被称为“元数据”)中学习，当面对新的环境与任务时，从知识库中找到与之最相似的元案例，从而快速找到合适的算法模型和超参数来解决新问题。由于这个优势，元学习已经被广泛用于解决分类、图像识别等领域的算法选择和超参数优化问题。然而，该框架存在的不足之处是：当面对一个在元知识库中没有出现过或与已有案例相似度很低的问题时，元学习的性能往往比较差。其次，需要大量的先验任务集，且假定任务之间都是独立同分布的。一旦任务发生动态变化导致不具有独立同分布的特性，该框架便不能及时适应这种变化。此外，该框架不能通过利用实际测试的性能信息改善学习模型，因此，很难处理复杂的数据和复杂的学习环境。
S.Shalev-Shwartz在其发表的论文“Online learning and online convex optimization”(Foundations and trends in Machine Learning，vol.4，no.2，pp.107-194，2011.)中提出了一种在 线学习算法框架。在线学习会按照顺序连续获得训练数据。它可以根据反馈的数据快速调整模型，提高了模型的准确性。然而，该框架存在的不足之处是：在线学习过程只专注于优化当前的问题。当新的任务到达时，由于没有利用之前的任务信息来获取模型的初始参数，会降低学习速度和精度。
C.Finn，A.Rajeswaran，S.Kakade and S.Levin等人在其发表的论文“Online Meta-Learning”(Proc.of the 36 th International Conference on Machine Learning(ICML-19)，pp.1920-1930)中将元学习与在线学习相结合，提出了在线元学习框架，它利用以前的经验获取先验知识，能够适应当前的任务。然而，该框架存在的不足之处是：与元学习类似，在线元学习只能适应经验中存在的任务变化，无法适应全新的任务和环境。此外，当元知识库中存在不良的训练样本时，该框架的学习性能会显著降低。
G.Zeng，Y.Chen，B.Cui and S.Yu等人在其发表的论文“Continual learning of context-dependent processing in neural networks”(Nat.Mach.Intell.，vol.1，pp.364-372，2019.)中提出了对神经网络进行连续学习的方法，可以逐渐学习现实世界中的种种规律。然而，该框架存在的不足之处是：如果训练数据不够精准，持续学习的性能也会显著下降。
华北理工大学提出的专利申请“具有脑认知机理的发育自动机及其学习方法”(申请号CN201510628233.0申请公布号CN105205533B)中公开了一种融合脑认知机理的发育自动机。该方法为系统自主发育过程提供了一种泛化能力强，适用范围广的数学模型，并将感觉运动系统与内在动机机制相结合，提高系统的自学习与自适应能力。然而，该框架存在的不足之处是：没有自我评估的功能，无法纠正因错误信号导致的非最优行为。
发明内容
本发明针对上述现有技术的不足，提出一种基于大脑机理的认知学习方法。
为实现上述技术目的，本发明采取的技术方案为：
一种基于大脑机理的认知学习方法，其特征是：包括如下步骤：
(1)在线学习：
(1a)接收来自外界的动态环境和动态任务；
(1b)认知特征提取模块提取出动态环境和动态任务的特征，并将其传给认知控制模块；
(1c)认知控制模块在步骤(1b)中得到的特征与合适的算法和超参数之间建立映射关系，以获得合适的算法类型和超参数组合；
(1d)认知控制模块从记忆模块的算法和超参数库调用步骤(1c)中选择的具体算法类型和超参数组合，并将其传给学习网络；
(1e)学习网络模块执行步骤(1d)中调用的算法和超参数，在步骤(1b)提取的动态环境和动态任务中执行超参数和算法，获得相应的学习结果，并将其输出到外界的动态环境中；
(3)离线自学习：
(2a)将在线学习步骤(1b)中提取的动态环境和动态任务的原始数据存储至记忆模块中的数据库，并将动态环境和任务的特征、步骤(1c)中选择的算法和超参数及步骤(1e)中得到的学习结果作为新的认知案例，存储至记忆模块中的认知案例库；
(2b)认知案例库采样认知案例；
(2c)认知控制模块对步骤(2b)中案例的特征与合适的算法和超参数之间建立映射关系，以获得合适的算法类型和超参数组合；
(2d)认知控制模块从记忆模块的算法和超参数库调用步骤(2c)中选择的具体算法类型和超参数组合，并将其传给学习网络；
(2e)学习网络基于步骤(2b)中的动态环境和任务执行由步骤(2d)中获得的算法和超参数，获得相应的学习结果，并将其传给认知评估模块；
(2f)认知案例库将步骤(2b)该案例过去的最佳学习结果传给认知评估模块；
(2g)认知评估模块比较步骤(2e)中该案例的新学习结果和步骤(2f)中该案例的历史最佳学习结果，将取得更优学习结果的算法和超参数以及该学习结果作为更新后的历史最佳算法和超参数、学习性能，传给认知案例库，更新步骤(2b)中的认知案例；
(2h)根据更新后的认知案例库重新训练认知控制模块。
为优化上述技术方案，采取的具体措施还包括：
步骤(1a)接收来的自外界的动态环境和动态任务为[e，x]＝d∈D，其中，e表示描述动态环境的数据，x表示描述动态任务的数据，d表示由这两者组成的向量，D为外界数据库，∈为属于号，表示动态环境和动态任务来源于外界数据库。
步骤(1b)认知特征提取模块提取出动态环境和动态任务的特征为f(e，x)，并将其传给认知控制模块。
步骤(1c)所述的采用认知控制模块，在步骤(1b)中得到的特征与合适的算法和超参数之间建立映射关系S[f(e，x)]，以获得合适的算法类型a∈A和超参数组合λ∈Λ，其中，A表示算法库，a表示其中是一种算法，A表示超参数库，λ表示其中的一种超参数组合，∈为属于号，表示动态环境和动态任务来源于外界数据库。
步骤(1e)所述的采用学习网络模块，执行步骤(1d)中调用的算法和超参数，在环境 和任务d中执行超参数为λ的算法a，获得相应的学习结果y＝a[λ，d]，并将其输出到外界的动态环境中。
步骤(2a)所述的采用认知案例库，将在线学习步骤(1b)中提取的动态环境和任务的原始数据[e，x]存储至记忆模块中的数据库
Figure PCTCN2021129297-appb-000001
并将动态环境和任务的特征f(e，x)、步骤(1c)中选择的算法和超参数(a，λ)及步骤(1e)中得到的学习结果y作为新的认知案例[f(e，x)，(a best，λ best)，y best]，存储至记忆模块中的认知案例库。
步骤(2 c)所述的采用认知控制模块，对步骤(2b)中案例的特征与合适的算法和超参数之间建立映射关系
Figure PCTCN2021129297-appb-000002
以获得合适的算法类型
Figure PCTCN2021129297-appb-000003
和超参数组合
Figure PCTCN2021129297-appb-000004
步骤(2e)所述的采用学习网络，基于步骤(2b)中的动态环境和任务
Figure PCTCN2021129297-appb-000005
执行由步骤(2d)中获得的算法和超参数，获得相应的学习结果
Figure PCTCN2021129297-appb-000006
并将其传给认知评估模块。
步骤(2g)所述的采用认知评估模块，比较步骤(2e)中该案例的新学习结果
Figure PCTCN2021129297-appb-000007
和步骤(2f)中该案例的历史最佳学习结果
Figure PCTCN2021129297-appb-000008
将取得更优学习结果的算法和超参数以及该学习结果作为更新后的历史最佳算法和超参数、学习性能，传给认知案例库，更新步骤(2b)中的认知案例
Figure PCTCN2021129297-appb-000009
根据人和灵长类动物的大脑机制，通过模拟人的认知机制和行为以改善智能的机器学习方法，是当前人工智能领域的重要创新点。研究表明，由于大脑中的感觉皮层，前额叶皮层，前运动皮层，前扣带皮层和记忆中心等区域相互协作，人和灵长类动物能够制定和完成复杂的计划来指导行为；能够根据前额叶皮层的活动，在环境和任务变化时灵活快速地改变自己的行为以适应动态变化；能够根据反馈信息修正错误行为并做出更好的决定。这些特征和活动都是由大脑认知机理决定的。
受到人和灵长类动物大脑认知机制的启发，我们发明了一种认知学习的机器学习框架。其根据环境和任务选择最合适的算法类型和参数，能在各种不同的环境和任务中获得良好的学习效果，并且能修正错误行为。该框架内部包括四个主要的模块，即认知特征提取，认知控制，学习网络以及记忆模块。其中记忆模块内部包含数据库，认知案例库，算法和超参数库，分别存储动态环境和任务的数据，认知案例以及具体的算法和超参数值。针对动态的环境和任务，可以灵活地选择出最合适的算法类型和超参数组合。同时通过“良币驱逐劣币”， 用正确标记数据纠正错误标记数据，实现对训练数据的鲁棒性。
本发明与现有技术相比具有以下优点：
第一，由于本发明借鉴大脑机理引入认知控制模块和记忆模块，使得可以调用存储在记忆模块中的认知案例学习结果，由认知控制模块更新环境和任务特征与应选算法和超参数之间的匹配关系，以此方式来不断提高学习能力，选择到更合适的算法和参数。克服了现有框架一旦训练完毕，学习性能便不会再提升的不足，使得本发明拥有自学习的能力。
第二，由于本发明借鉴大脑机理引入认知特征提取模块，使得可以在环境或任务发生变化时，感知到这一变化并提取动态环境和任务的新特征。认知控制模块根据环境和任务特征与算法和超参数选择之间的匹配关系，根据新的特征改变算法类型和超参数组合的选择。克服了现有框架因算法和超参数固定而不能适应动态变化的场景，在新场景中表现不佳的不足。使得本发明可以为不同场景选择出最合适的算法和超参数，具有适应动态环境和任务的能力。
第三，由于本发明借鉴大脑机理引入认知评估模块，使得可以在离线自学习时，对所选算法类型和超参数在某个场景中的性能进行评估，将当前的学习结果与认知案例中的结果进行比较，纠正错误标记的训练样本。克服了现有框架极度依赖训练样本质量的不足，使得本发明可以具有对错误标记训练样本的鲁棒性，减少不佳的训练样本，提升学习性能。使得本发明具有“良币驱逐劣币”的能力。
附图说明
图1是本发明的流程图；
图2(a)至图2(d)是采用本发明和现有技术在测试次序数不断增加时的算法以及超参数选择准确率、图像识别准确率对比图；
图3(a)和图3(b)是采用本发明和现有技术在环境和任务不断动态转换时的算法以及超参数选择准确率对比图；
图4(a)至图4(d)是采用本发明和现有技术在不同错误标记训练样本量下的算法以及超参数选择准确率、图像识别准确率对比图。
具体实施方式
以下对本发明的实施例作进一步详细描述。
下面结合附图对发明做进一步描述。
结合附图1对本发明方法的具体步骤描述如下。
1.在线学习：
在线学习过程对应脑认知机制的执行控制过程。执行控制是个体在以目标为导向的行为过程中动态灵活地调节多个认知子系统活动的复杂认知过程。它对外部任务和数据进行两种操作，即制定计划来指导行为，和在发生意外变化时进行快速调节。这种机制使灵长类动物能够在动态环境中快速切换行为。
步骤1.1，接收来自外界的动态环境和动态任务[e，x]＝d∈D。其中，e表示描述动态环境的数据，x表示描述动态任务的数据，d表示由这两者组成的向量，D为外界数据库，∈为属于号，表示动态环境和动态任务来源于外界数据库。
步骤1.2，认知特征提取模块对应于大脑皮层中的感觉皮层，提取出数据的特征f(e，x)，并将其传给认知控制模块。
步骤1.3，认知控制模块对应于大脑皮层中的前额叶皮层，整合感觉皮层的特征，并基于特征信息和相关经验知识进行认知控制，获取刺激与反应之间的关联信息。在步骤1.2中得到的特征与合适的算法和超参数之间建立映射关系S[f(e，x)]，以获得合适的算法类型a∈A和超参数组合λ∈Λ。其中，A表示算法库，a表示其中的一种算法，A表示超参数库，λ表示其中的一种超参数组合，∈为属于号，表示算法来源于算法库，超参数组合来源于超参数库。
步骤1.4，记忆模块对应于大脑中的记忆中心，认知控制模块从记忆模块的算法库A和超参数库Λ调用步骤1.3中选择的具体算法类型和超参数组合，并将其传给学习网络。
步骤1.5，学习网络模块对应于前运动皮层，根据刺激与反应的关联信息得到具体的运动计划，执行步骤1.4中调用的算法和超参数，在环境和任务d中执行超参数为λ的算法a获得相应的学习结果y＝a[λ，d]，并将其输出到外界的动态环境中。
2.离线自学习：
离线自学习过程对应脑认知机制的内省过程。由于执行控制过程是一个快速的过程，可能导致不适当的运动反应。通过内省存储在内部记忆中导致不适当运动反应的事件，内省过程可以执行更适当的运动反应。内省过程是一种自发的心理活动，与当前的任务或感知的环境无关。
步骤2.1，记忆模块对应于大脑中的记忆中心，将在线学习步骤1.2中提取的动态环境和任务的原始数据[e，x]存储至记忆模块中的数据库
Figure PCTCN2021129297-appb-000010
并将动态环境和任务的特征f(e，x)、步骤1.3中选择的算法和超参数(a，λ)及步骤1.5中得到的学习结果y作为新的认知案例 [f(e，x)，(a best，λ best)，y best]，存储至记忆模块中的认知案例库。
步骤2.2，认知案例库采样认知案例
Figure PCTCN2021129297-appb-000011
步骤2.3，认知控制模块对应于大脑皮层中的前额叶皮层，整合感觉皮层的特征，并基于特征信息和相关经验知识进行认知控制，获取刺激与反应之间的关联信息，对步骤2.2中案例的特征与合适的算法和超参数之间建立映射关系
Figure PCTCN2021129297-appb-000012
以获得合适的算法类型
Figure PCTCN2021129297-appb-000013
和超参数组合
Figure PCTCN2021129297-appb-000014
步骤2.4，认知控制模块从记忆模块的算法和超参数库调用步骤2.3中选择的具体算法类型和超参数组合，并将其传给学习网络。
步骤2.5，学习网络模块对应于前运动皮层，根据刺激与反应的关联信息得到具体的运动计划，基于步骤2.2中的动态环境和任务
Figure PCTCN2021129297-appb-000015
执行由步骤2.4中获得的算法和超参数，获得相应的学习结果
Figure PCTCN2021129297-appb-000016
并将其传给认知评估模块。
步骤2.6，记忆模块中的认知案例库将步骤2.2中该案例过去的最佳学习结果
Figure PCTCN2021129297-appb-000017
传给认知评估模块。
步骤2.7，认知评估模块对应于大脑皮层中的前扣带皮层，监控前运动皮层的冲突反应，比较步骤2.5中该案例的新学习结果
Figure PCTCN2021129297-appb-000018
和步骤2.6中该案例的历史最佳学习结果
Figure PCTCN2021129297-appb-000019
然后，将冲突信息反馈给前额叶皮层，即令取得更优学习结果的算法和超参数以及该学习结果作为更新后的历史最佳算法和超参数、学习性能，传给认知案例库，更新步骤2.2中的认知案例
Figure PCTCN2021129297-appb-000020
若采样的案例的历史最佳学习结果y′ *优于新学习结果y′，则不更新。否则，按下式更新：
Figure PCTCN2021129297-appb-000021
步骤2.8，根据更新后的认知案例库重新训练认知控制模块。
下面结合仿真实验对本发明的效果做进一步说明。
1、仿真条件：
本发明的仿真实验在MATLAB R2016a软件下进行的。本发明的实例为图像分类任务。
实例共有5种不同的环境即5个数据集，包括公开数据集MNIST，FashionMNIST，Cifar-10。另外，还有一个取自ImageNet的子数据集，从ImageNet中取5类，每一类有4500个训练样本。和一个取自MNIST的子数据集，每一类仅取1000个训练样本。分别命名为mini-ImageNet 和mini-MNIST。实例共有2个任务，即图像分类精度百分比需求和图像分类所需时间需求。
实例的认知特征包含五个数据集特征：数据集中图像的训练样本总数，数据集中图像分类类别个数，数据集中图像宽度(图像被调整为长度与宽度相等)，数据集中图像色彩特征，图像分类难度；与两个性能需求特征：图像分类准确率的需求；对图像分类完成时间的需求。
实例的认知控制模块，针对算法选择采取径向基函数神经网络(RBFNN)的结构，针对超参数选择采取BP神经网络的结构。
实例的算法库中包含支持向量机，Adaboosting，随机森林，决策树，深度神经网络五种不同的分类算法。超参数库中包含这些算法各自的重要超参数。
2、仿真内容：
图2(a)至图2(d)是采用本发明和现有技术在测试次序数不断增加时的算法以及超参数选择准确率、图像识别准确率对比图；
图3(a)和图3(b)是采用本发明和现有技术在环境和任务不断动态转换时的算法以及超参数选择准确率对比图；
图4(a)至图4(d)是采用本发明和现有技术在不同错误标记训练样本量下的算法以及超参数选择准确率、图像识别准确率对比图。
参照附图2(a)至图2(d)，采用本发明和现有技术在测试次序数不断增加时的算法以及深度神经网络的超参数选择准确率、图像识别准确率的对比结果的曲线图中可以看出，现有技术中，在首次测试前得到固定的元示例库，随着测试次数增加，其算法以及超参数选择准确率、图像识别准确率虽然略有波动，但几乎维持稳定。而本方法认知案例空间的大小可以不断扩大，通过吸收测试案例作为训练样本对神经网络进行再训练，这些性能都会逐渐得到改善。结合附图2(a)和附图2(b)，以及附图2(c)和附图2(d)，还可以看出随着算法以及超参数选择准确率的提高，图像识别准确率也得到了提高。
附图2(a)和附图2(b)分别是采用本发明和现有技术在测试次序数不断增加时的算法选择准确率和基于此算法的图像分类准确率对比图。算法选择精度评估是通过检查选择的算法是否为标记的最佳算法。图像分类精度评价是通过检查根据已选算法及其最佳超参数执行图像分类的分类结果是否正确。横坐标表示测试次序数(次)，纵坐标表示算法选择/图像分类准确率(百分比)。以红色方格标示的折线表示采用本发明方法的算法选择/图像分类准确率，由于本发明可以存储已被测试的案例至认知案例空间，六次测试时，其针对每个数据集的实际训练样本数分别为30、50、100、150、250、400。以黑色圆圈标示的折线表示现有技术从具有针对每个数据集30个元案例的库实现算法选择的算法选择/图像分类准确率，以蓝色正三角标示的折线表示现有技术从具有针对每个数据集50个元案例的库实现算法选择的算法 选择/图像分类准确率，以绿色倒三角标示的折线表示现有技术从具有针对每个数据集250个元案例的库实现算法选择的算法选择/图像分类准确率。通过比较两种方法得到的算法选择/图像分类精度变化情况，可以看出，当本发明的训练样本数等于现有方法的元案例个数时，两种方法性能几乎一致。但是采用本发明方法得到的算法选择/图像分类准确率随着测试次数增加而提高，采用现有技术得到的算法选择/图像分类准确率却几乎不变。
附图2(c)和附图2(d)分别是采用本发明和现有技术在测试次序数不断增加时的深度神经网络超参数选择准确率和基于此超参数的深度神经网络图像分类准确率对比图。深度神经网络超参数选择精度评估是通过检查选择的超参数是否为标记的最佳超参数。横坐标表示测试次序数(次)，纵坐标表示超参数选择/图像分类准确率(百分比)。以红色方格标示的折线表示采用本发明方法的超参数选择/图像分类准确率，由于本发明可以存储已被测试的案例至认知案例空间，六次测试时，其针对每个数据集的训练样本数分别为50、100、150、200、250、300。以黑色圆圈标示的折线表示现有技术从具有针对每个数据集50个元案例的库实现超参数选择的超参数选择/图像分类准确率，以蓝色正三角标示的折线表示现有技术从具有针对每个数据集100个元案例的库实现超参数选择的超参数选择/图像分类准确率，以绿色倒三角标示的折线表示现有技术从具有针对每个数据集250个元案例的库实现超参数选择的超参数选择/图像分类准确率。通过比较两种方法得到的超参数选择选择/图像分类精度变化情况，可以看出，当本发明的训练样本数等于现有方法的元案例个数时，两种方法性能几乎一致。但是采用本发明方法得到的超参数选择/图像分类准确率随着测试次数增加而提高，采用现有技术得到的超参数选择/图像分类准确率却几乎不变。
参照附图3(a)和图3(b)，采用本发明和现有技术在环境和任务不断动态转换时的算法以及超参数选择准确率对比结果所得的曲线图中，可以看出，不管环境和任务如何变化，本发明都能很快地适应，只需要很少的次数就能达到较好的精度。其中，动态环境指数据集的变化，动态任务指性能需求的变化。完成时间优先是指标记算法是完成时间最少且满足精度要求的算法。精度优先是指标记算法是图像分类精度最高的算法，而完成时间有相对宽松的约束。最高完成时间的限制为6000秒，我们将完成时间低于3000秒的案例看作时间优先，完成时间高于3000秒的案例看作精度优先。在仿真中，环境的变化导致图像数据集特征的变化，任务的变化导致图像分类精度和完成时间的需求变化。现有方法只能在已训练样本上表现出良好性能，一旦环境和任务变化，由于元案例库固定而不能适应，性能持续较低。
附图3(a)是采用本发明和现有技术在环境和任务不断动态转换时的算法选择准确率对比图。横坐标表示测试次序数(次)，纵坐标表示算法选择准确率(百分比)。以红色方格标示的折线表示采用本发明方法的算法选择准确率。以蓝色三角标示的折线表示现有技术从具 有针对mini-ImageNet 50个时间优先的元案例的库实现算法选择的算法选择准确率。通过比较两种方法得到的算法选择精度变化情况，可以看出，当一开始第1～2次测试在mini-ImageNet上选择时间优先的算法时，两种方法的选择精度几乎一致。但是当第3～7次测试任务变为精度优先，本发明方法与现有方法一开始的算法选择准确率都降至很低，但是随着测试次数增加本发明方法迅速提高，采用现有技术得到的算法选择准确率却一直较低。当第8～10次测试任务再次变为时间优先，本发明方法的性能已经逐渐优于现有方法。当第11～15次测试数据集变化为cifar-10，任务为完成时间优先时，本发明方法与现有方法一开始的算法选择准确率都降至很低，但是随着测试次数增加本发明方法迅速提高，采用现有技术得到的算法选择准确率却一直较低。
附图3(b)是采用本发明和现有技术在环境和任务不断动态转换时的超参数选择准确率对比图。横坐标表示测试次序数(次)，纵坐标表示超参数选择准确率(百分比)。以红色方格标示的折线表示采用本发明方法的超参数选择准确率。以蓝色三角标示的折线表示现有技术从具有针对mini-ImageNet 50个时间优先的元案例的库实现超参数选择的超参数选择准确率。通过比较两种方法得到的超参数选择精度变化情况，可以看出，当一开始第1～2次测试在mini-ImageNet上选择时间优先的超参数时，两种方法的选择精度几乎一致。但是当第3～11次测试任务变为精度优先，本发明方法与现有方法一开始的超参数选择准确率都降至很低，但是随着测试次数增加本发明方法迅速提高，采用现有技术得到的超参数选择准确率却一直较低。当第12～13次测试任务再次变为时间优先，本发明方法的性能已经逐渐优于现有方法。当第14～22次测试数据集变化为cifar-10，任务为完成时间优先时，本发明方法与现有方法一开始的超参数选择准确率都降至很低，但是随着测试次数增加本发明方法迅速提高，采用现有技术得到的超参数选择准确率却一直较低。
参照附图4(a)至图4(d)，采用本发明和现有技术在不同数量错误标记训练样本下的算法以及超参数选择准确率、图像识别准确率的对比结果的曲线图中可以看出，不管错误标记的样本比例是多少，通过评估案例并纠正错误案例，本发明中算法以及超参数选择准确率、图像识别准确率性能会逐渐得到改善。现有技术中，在首次测试前得到固定的元示例库，错误标记样本将一直存储在元案例库中，持续影响性能。劣币指的是认知案例空间中的错误标签案例。劣币率用错误标记案例的比例来表示。算法或超参数选择的训练样本可能贴错了标签，此时面对测试样本很可能作出一个错误的选择。结合附图4(a)和附图4(b)，以及附图4(c)和附图4(d)，可以看出随着算法以及超参数选择准确率的提高，图像识别准确率也得到了提高。
附图4(a)和(b)分别是采用本发明和现有技术在测试次数不断增加时的算法选择准确率/基于算法选择的图像分类准确率对比图。横坐标表示测试次序数(次)，纵坐标表示算法选 择/图像分类准确率(百分比)。以红色方格标示的折线表示劣币率为10％时采用本发明方法的算法选择/图像分类准确率，以橙色圆形标示的折线表示劣币率为30％时采用本发明方法的算法选择/图像分类准确率。以蓝色正三角标示的折线表示劣币率为10％时现有技术的算法选择/图像分类准确率，以绿色倒三角标示的折线表示劣币率为30％时现有技术的算法选择/图像分类准确率。通过比较两种方法得到的算法选择/图像分类精度变化情况，可以看出，本发明所取得的算法选择/图像分类精度优于现有方法。本发明与现有方法一开始都受到劣币的影响，但是本发明的性能会逐渐提升。这是因为一方面，更好的环境和任务特征与算法和超参数间的匹配关系存储在认知案例空间，错误标记样本量的比例降低。另一方面，在本发明中认知评估模块可以将当前的学习结果与之前的学习结果进行比较，并将评价结果反馈给认知控制模块。而采用现有技术始终受到错误标记样本的影响，得到的算法选/图像分类择准确率却几乎不变。
附图4(c)和(d)分别是采用本发明和现有技术在测试次数不断增加时的超参数选择准确率/深度神经网络中基于超参数选择的图像分类准确率对比图。横坐标表示测试次序数(次)，纵坐标表示超参数选择/图像分类准确率(百分比)。以红色方格标示的折线表示劣币率为10％时采用本发明方法的超参数选择/图像分类准确率，以橙色圆形标示的折线表示劣币率为30％时采用本发明方法的超参数选择/图像分类准确率。以蓝色正三角标示的折线表示劣币率为10％时现有技术的超参数选择/图像分类准确率，以绿色倒三角标示的折线表示劣币率为30％时现有技术的超参数选择/图像分类准确率。通过比较两种方法得到的精度变化情况，可以看出，本发明所取得的超参数选择/图像分类精度优于现有方法。本发明与现有方法一开始都受到劣币的影响，但是本发明的性能会逐渐提升。这是因为一方面，更好的环境和任务特征与算法和超参数间的匹配关系存储在认知案例空间，错误标记样本量的比例降低。另一方面，在本发明中认知评估模块可以将当前的学习结果与之前的学习结果进行比较，并将评价结果反馈给认知控制模块。而采用现有技术始终受到错误标记样本的影响，得到的超参数选择/图像分类择准确率几乎不变。
综合上述仿真结果和分析，本发明所提出的基于大脑机理的认知学习框架与方法，具有自学习、适应动态变化的环境和任务、良币驱除劣币的能力，这使得本发明在实际中能更好的得到应用。
以上仅是本发明的优选实施方式，本发明的保护范围并不仅局限于上述实施例，凡属于本发明思路下的技术方案均属于本发明的保护范围。应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明原理前提下的若干改进和润饰，应视为本发明的保护范围。

Claims (9)

  1. 一种基于大脑机理的认知学习方法，其特征是：包括如下步骤：
    (1)在线学习：
    (1a)接收来自外界的动态环境和动态任务；
    (1b)认知特征提取模块提取出动态环境和动态任务的特征，并将其传给认知控制模块；
    (1c)认知控制模块在步骤(1b)中得到的特征与合适的算法和超参数之间建立映射关系，以获得合适的算法类型和超参数组合；
    (1d)认知控制模块从记忆模块的算法和超参数库调用步骤(1c)中选择的具体算法类型和超参数组合，并将其传给学习网络；
    (1e)学习网络模块执行步骤(1d)中调用的算法和超参数，在步骤(1b)提取的动态环境和动态任务中执行超参数和算法，获得相应的学习结果，并将其输出到外界的动态环境中；
    (2)离线自学习：
    (2a)将在线学习步骤(1b)中提取的动态环境和动态任务的原始数据存储至记忆模块中的数据库，并将动态环境和任务的特征、步骤(1c)中选择的算法和超参数及步骤(1e)中得到的学习结果作为新的认知案例，存储至记忆模块中的认知案例库；
    (2b)认知案例库采样认知案例；
    (2c)认知控制模块对步骤(2b)中案例的特征与合适的算法和超参数之间建立映射关系，以获得合适的算法类型和超参数组合；
    (2d)认知控制模块从记忆模块的算法和超参数库调用步骤(2c)中选择的具体算法类型和超参数组合，并将其传给学习网络；
    (2e)学习网络基于步骤(2b)中的动态环境和任务执行由步骤(2d)中获得的算法和超参数，获得相应的学习结果，并将其传给认知评估模块；
    (2f)认知案例库将步骤(2b)该案例过去的最佳学习结果传给认知评估模块；
    (2g)认知评估模块比较步骤(2e)中该案例的新学习结果和步骤(2f)中该案例的历史最佳学习结果，将取得更优学习结果的算法和超参数以及该学习结果作为更新后的历史最佳算法和超参数、学习性能，传给认知案例库，更新步骤(2b)中的认知案例；
    (2h)根据更新后的认知案例库重新训练认知控制模块。
  2. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(1a)接收来的自外界的动态环境和动态任务为[e，x]＝d∈D，其中，e表示描述动态环境的数据，x表示描述动态任务的数据，d表示由这两者组成的向量，D为外界数据库，∈为属于号，表示动 态环境和动态任务来源于外界数据库。
  3. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(1b)认知特征提取模块提取出动态环境和动态任务的特征为f(e，x)，并将其传给认知控制模块。
  4. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(1c)所述的采用认知控制模块，在步骤(1b)中得到的特征与合适的算法和超参数之间建立映射关系S[f(e，x)]，以获得合适的算法类型a∈A和超参数组合λ∈Λ，其中，A表示算法库，a表示其中是一种算法，Λ表示超参数库，λ表示其中的一种超参数组合，∈为属于号，表示动态环境和动态任务来源于外界数据库。
  5. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(1e)所述的采用学习网络模块，执行步骤(1d)中调用的算法和超参数，在环境和任务d中执行超参数为λ的算法a，获得相应的学习结果y＝a[λ，d]，并将其输出到外界的动态环境中。
  6. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(2a)所述的采用认知案例库，将在线学习步骤(1b)中提取的动态环境和任务的原始数据[e，x]存储至记忆模块中的数据库
    Figure PCTCN2021129297-appb-100001
    并将动态环境和任务的特征f(e，x)、步骤(1c)中选择的算法和超参数(a，λ)及步骤(1e)中得到的学习结果y作为新的认知案例[f(e，x)，(a best，λ best)，y best]，存储至记忆模块中的认知案例库。
  7. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(2c)所述的采用认知控制模块，对步骤(2b)中案例的特征与合适的算法和超参数之间建立映射关系
    Figure PCTCN2021129297-appb-100002
    以获得合适的算法类型
    Figure PCTCN2021129297-appb-100003
    和超参数组合
    Figure PCTCN2021129297-appb-100004
  8. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(2e)所述的采用学习网络，基于步骤(2b)中的动态环境和任务
    Figure PCTCN2021129297-appb-100005
    执行由步骤(2d)中获得的算法和超参数，获得相应的学习结果
    Figure PCTCN2021129297-appb-100006
    并将其传给认知评估模块。
  9. 根据权利要求1所述的一种基于大脑机理的认知学习方法，其特征是：步骤(2g)所述的采用认知评估模块，比较步骤(2e)中该案例的新学习结果
    Figure PCTCN2021129297-appb-100007
    和步骤(2f)中该案例的历史最佳学习结果
    Figure PCTCN2021129297-appb-100008
    将取得更优学习结果的算法和超参数以及该学习结果作为更新后的历史最佳算法和超参数、学习性能，传给认知案例库，更新步骤(2b)中的认知案例
    Figure PCTCN2021129297-appb-100009
PCT/CN2021/129297 2021-05-25 2021-11-08 一种基于大脑机理的认知学习方法 Ceased WO2022247151A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/786,564 US11948092B2 (en) 2021-05-25 2021-11-08 Brain-inspired cognitive learning method

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110569553.9A CN113255765B (zh) 2021-05-25 2021-05-25 一种基于大脑机理的认知学习方法
CN202110569553.9 2021-05-25

Publications (1)

Publication Number Publication Date
WO2022247151A1 true WO2022247151A1 (zh) 2022-12-01

Family

ID=77184336

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/129297 Ceased WO2022247151A1 (zh) 2021-05-25 2021-11-08 一种基于大脑机理的认知学习方法

Country Status (3)

Country Link
US (1) US11948092B2 (zh)
CN (1) CN113255765B (zh)
WO (1) WO2022247151A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113255765B (zh) 2021-05-25 2024-03-19 南京航空航天大学 一种基于大脑机理的认知学习方法
CN116542286B (zh) * 2022-01-21 2026-03-27 北京灵汐科技有限公司 模型超参数取值方法及装置、处理核、设备、芯片和介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170091664A1 (en) * 2015-09-29 2017-03-30 Cognitive Scale, Inc. Cognitive Machine Learning System
CN109447140A (zh) * 2018-10-19 2019-03-08 广州四十五度科技有限公司 一种基于神经网络深度学习的图像识别并推荐认知的方法
CN111714118A (zh) * 2020-06-08 2020-09-29 北京航天自动控制研究所 一种基于集成学习的脑认知模型融合方法
CN113255765A (zh) * 2021-05-25 2021-08-13 南京航空航天大学 一种基于大脑机理的认知学习方法

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5724987A (en) * 1991-09-26 1998-03-10 Sam Technology, Inc. Neurocognitive adaptive computer-aided training method and system
CN101466111B (zh) * 2009-01-13 2010-11-17 中国人民解放军理工大学通信工程学院 基于政策规划约束q学习的动态频谱接入方法
CN105205533B (zh) 2015-09-29 2018-01-05 华北理工大学 具有脑认知机理的发育自动机及其学习方法
CN108809456B (zh) * 2018-07-04 2020-09-18 天津大学 一种基于改进强化学习的集中式认知无线电频谱分配方法
US11663523B2 (en) * 2019-09-14 2023-05-30 Oracle International Corporation Machine learning (ML) infrastructure techniques
CN111210023B (zh) * 2020-01-13 2023-04-11 哈尔滨工业大学 数据集分类学习算法自动选择系统及方法
CN111461294B (zh) * 2020-03-16 2022-10-11 中国人民解放军空军工程大学 面向动态博弈的智能飞行器类脑认知学习方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170091664A1 (en) * 2015-09-29 2017-03-30 Cognitive Scale, Inc. Cognitive Machine Learning System
CN109447140A (zh) * 2018-10-19 2019-03-08 广州四十五度科技有限公司 一种基于神经网络深度学习的图像识别并推荐认知的方法
CN111714118A (zh) * 2020-06-08 2020-09-29 北京航天自动控制研究所 一种基于集成学习的脑认知模型融合方法
CN113255765A (zh) * 2021-05-25 2021-08-13 南京航空航天大学 一种基于大脑机理的认知学习方法

Also Published As

Publication number Publication date
US20230133628A1 (en) 2023-05-04
CN113255765A (zh) 2021-08-13
CN113255765B (zh) 2024-03-19
US11948092B2 (en) 2024-04-02

Similar Documents

Publication Publication Date Title
Blauch et al. Computational insights into human perceptual expertise for familiar and unfamiliar face recognition
Zhang et al. An novel end-to-end network for automatic student engagement recognition
Jagadeesh et al. Facial expression recognition of online learners from real-time videos using a novel deep learning model
Cheng et al. Disentangled prompt representation for domain generalization
CN117058752B (zh) 一种基于改进YOLOv7的学生课堂行为检测方法
CN109145871A (zh) 心理行为识别方法、装置与存储介质
WO2022247151A1 (zh) 一种基于大脑机理的认知学习方法
CN114881155B (zh) 基于深度迁移学习的水果图像分类方法
CN118411275A (zh) 一种基于大数据的教学系统
Song Emotional recognition and feedback of students in English e-learning based on computer vision and face recognition algorithms
CN113762402B (zh) 基于小样本持续学习的图像分类方法、装置及存储介质
CN116778223B (zh) 一种基于零标注的多类别目标域数据集标注方法及系统
Kousalya et al. Prediction of Best Optimizer for Facial Expression Detection using Convolutional Neural Network
Prakash et al. SmartLeafNet: An Automated Deep Learning based Efficient Plant Disease Detection using RNN-GRU Model
Kumar et al. Detection of Mango Leaves Disease using CSUBW based Residual Attention with EfficientNet
Xu et al. Automatic cataract grading with visual-semantic interpretability
Jesupriya et al. Deep Learning-Based Transfer Learning with MobileNetV2 for Crop Disease Detection
Yung-Hui et al. Clinically applicable deep learning for diagnosis of diabetic retinopathy
Topolski et al. Modification of the Principal Component Analysis Method Based on Feature Rotation by Class Centroids.
Lin Online Professional-Creative Fusion Music Major Students’ Classroom State Recognition Based on the Integration of DSC and LeNet-5 Models
Gbolagade et al. Predicting postgraduate performance using resample pre-process algorithm and artificial neural network
AU2021339829B2 (en) Automated aneuploidy screening using arbitrated ensembles
Ghadekar et al. Multimodal Deep Learning System for Nutritional Deficiency Assessment and Dietary Recommendation
Lin et al. GReFC-Net: an automated method for measuring structural features of spiral galaxies
Arafiyah et al. Performance Analysis of Convolutional Neural Network Architecture in Batik Design Classification for Mobile Application

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21942715

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21942715

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 21942715

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 28/06/2024)

122 Ep: pct application non-entry in european phase

Ref document number: 21942715

Country of ref document: EP

Kind code of ref document: A1