浙江大学人工智能研究所(以下简称AI)创建于1981年,是专概要
- 格式:doc
- 大小:22.00 KB
- 文档页数:2
近日,人工智能专业作为战略新兴产业受到关注,高考圈整理了目前人工智能全国排名前十的大学,供家长、考生了解。
中国科学院大学中国科学院的自动化研究所在人工智能领域的研究实力非常强大。
2017年5月,中国科学院大学成立人工智能技术学院。
这是我国人工智能技术领域较早的全面开展教学和科研工作的新型学院。
该学院就是由中国科学院自动化所牵头新成立的。
清华大学2018年6月28日,清华大学人工智能研究院在李兆基科技大楼揭牌成立。
由清华大学计算机系教授、中国科学院院士张钹出任首任院长。
清华大学的智能技术与系统国家重点实验室,称得上是国内在人工智能人才培养和科学研究的重要基地。
北京大学北大的信息科学技术学院下设的智能科学与技术专业由北大学数学系、计算机系、电子学系等10个系(所)于1985年成立,主要从事机器感知、智能机器人、智能信息处理和机器学习等交叉学科的研究和教学。
浙江大学浙江大学计算机学院下设的人工智能研究所是中国设立最早的人工智能研究机构之一。
在1978年就开始了人工智能领域的科学研究和人才培养,在1982年创建了人工智能研究室(1987年升级为研究所)。
校长吴朝晖院士、中国工程院原常务副院长潘云鹤院士都是目前学校人工智能研究领域的著名专家学者。
到现在,人工智能进入大数据阶段,浙大在计算机视觉领域已经建立了相当大的优势。
哈尔滨工业大学哈工大的可谓王牌工科院校,在全国工科高校实力排行榜中位居第二!仅次于清华!在全国高校学科评估中,哈工大的计算机科学与技术学科位列全国第4名,是国家重点一级学科,并进入ESI全球前1%的研究机构行列。
中国科学技术大学中国科学院的自动化研究所在工业自动化、智能设备控制、模式识别、智能信息处理等领域的成就享誉国内外,号称中国人工智能领域的黄埔军校。
另外,中国科学科技大学的科学与技术学院是教育部和国家计委首批批准的国家示范软件学院。
该学院的科研力量主要集中在高性能计算、智能计算与应用、网络计算与可信计算、先进计算机系统四大领域,实力非常强悍。
浙江大学2016 –2017 学年春夏学期《Artificial Intelligence》课程期末考试试卷课程号: 21191890 ,开课学院:_计算机科学与技术学院_考试试卷:A卷、B卷(请在选定项上打√)考试形式:闭、开卷(请在选定项上打√),允许带___________入场考试日期: 2017 年 6 月 25 日,考试时间: 120 分钟诚信考试,沉着应考,杜绝违纪。
考生姓名:_____________学号:_____________所属院系:_____________1.Fill in the blanks (30 points)1)Two common structures are used, namely first-in-first-out (FIFO queue) and last-in-first-out (LIFO queue). The breadth-first-search uses a __________queue, depth-first search uses a__________ queue.2)In alpha-beta pruning search, the algorithm maintains two values, alpha and beta, whichrepresents the maximum score that the maximizing player is assured of and the minimum score that the minimizing player is assured of respectively. At the beginning of alpha-beta search, alpha is set to __________ and beta is set to __________, i.e. both players start with their lowest possible score.3) A and B are two random variables. P(A) and P(B) are their probabilities. It is knownthat P(A)+P(B)=0.5, and P(A|B)P(B|A)=14. Then the P(A) is_______.4) In your 10-day vacation in Alaska, you kept the following log on the weather andwhether you saw a bear that day:(rain, bear) 1 day (¬rain, bear) 2 days (rain, ¬bear) 6 days (¬rain, ¬bear)1 daya) Compute the marginal probability P(bear ) = _______b) Compute the conditional probability P(¬bear|rain) = _______5) In the figure, the circles show a plot of a trainingdata set of 10 data points, the dash line shows the function f(x) used to generate the data and solid curve shows the higher order polynomial g(x) fitted to given 10 data points. The fitted curve passes exactly through each data point, so the value of RMS error E RMS = 0 and g(x) gives avery poor representation of the function f(x). This behavior is known as___________________ (please select over-fitting or under-fitting to fill in this blank).6) For a given likelihood function p(x n |θ), if we obtain a data set of observations X ={x 1,x 2,x 3} and these data points are independent and identically distributed (i.i.d.), then p (X |θ)=p(x 1,x 2,x 3|θ)= ____________.7) For multivariate Gaussian distribution N (x|μ, Σ) of the D dimensional input space x,we have __________ independent parameters for μ and Σ. If Σ is a diagonal matrix and 2σ∑=I , the number of total parameters reduces to __________.8) The Linear basis function models involve linear combinations of fixed nonlinearfunctions of the input variables. If given basis functions ϕ(x )=(ϕ0(x ),ϕ1(x ),ϕ2(x ))T, where ϕ0(x )=1 and the model parameters w =(w 0,w 1,w 2)T , then the linear basis function y (x , w ) = ____________.9)In general, a deep convolutional neural network consists of convolutional layer,pooling layer, fully-connected layer and classifier layer, the softmax is usuallyemployed at the __________ layer.10)Reinforcement learning mainly consists of policy, value function and model.A__________ maps a state to an action, and a value function is a prediction of future reward. In Q-value function, discount factor γ is usually used, the range of discount factor γ is __________.2. Multiple Choice (36 points, only one of the options is correct.)1)Consider three 2D points a = (0, 0), b = (0, 1), c = (1, 0). Run k-means with twoclusters. Let the initial cluster centers be (-1, 0), (0, 2). What clusters will k-means learn after one iteration? _____(A) {a}, {b, c}(B) {a, b}, {c}(C) {a, c}, {b}(D) none of the above2)The sigmoid function in a neural network is defined as g(x)=e x1+e x. There isanother commonly used activation function called the hyperbolic tangent function,which is defined as tanh(x)=e x−e−xe x+e−x. How are these two functions related?____________(A) tanℎ(x)=g(x)−1(B) tanℎ(x)=2g(x)−1(C) tanℎ(x)=g(2x)−1(D) tanℎ(x)=2g(2x)−13)Which nodes will be pruned along with their branches by alpha-beta pruning? _____(A) I (B) H, I (C) G, H, I (D) C, H, I4)Consider a 3-puzzle where, like in the usual 8-puzzle game, a tile canonly move to an adjacent empty space. Given the initial state which ofthe following state cannot be reached? _____(A) (B) (C) (D)5)Given two Gaussian distribution N(x|−1,1) and N(x|1,1), which of the followingformula is correct? ______________(A) N(0|−1,1)>N(0|1,1)(B) N(−1|−1,1)>N(−1|1,1)(C) N(0|−1,1)<N(0|1,1)(D) N(−1|−1,1)<N(−1|1,1)6)The Fisher’s criterion is defined to be _________(A)the separation of the projected class means.(B)the separation of the projected class variances.(C)the ratio of the between-class variance to the within-class variance.(D)the ratio of the within-class variance to the between-class variance.7)Suppose we have a data set {x1,…,x N} drawn from the mixture of two 2D Gaussians,which can be written as p(x)=0.5N(x|μ1,Σ1)+0.5N(x|μ2,Σ2). If Σ1=Σ2=σ2I in this model, which of the following figures is consistent with the distribution of data points p(x)? _______________(A) (B) (C) (D)8)Consider a polynomial curve fitting problem. If the fitted curve oscillates wildlythrough each point and achieve bad generalization by making accurate predictions for new data, we say this behavior is over-fitting. Which of the following methods cannot be used to control over-fitting? ________________(A)Use fewer training data(B)Add validation set, use Cross-validation(C)Add a regularization term to an error function(D)Use Bayesian approach with suitable prior9)AlexNet (one of popular multi-layer convolutional neural networks for imageclassification) is trained in a_________ setting, K-means clustering is employed in a _________ setting and Boosting for classification is implemented in a_________ setting , linear regression model for classification is realized in a _________ setting.(A) unsupervised, supervised, supervised, unsupervised(B) supervised, supervised, supervised, supervised(C) supervised, supervised, unsupervised, unsupervised(D) supervised, unsupervised, supervised, supervised10)In linear least square regression model, we can add a regularization term to an errorfunction (i.e., sum-of-squares) in order to control _________. The lasso regularizer will introduce____________ solution compared to quadratic regularizer.(A) over-fitting, dense (B) over-fitting, sparse(C) under-fitting, dense (D) under-fitting, sparse11)We can decompose the expected squared loss of one predict model as follows:expected loss = (bias)2 + variance + noiseIn general, one flexible model (i.e., under-fitting) model will introduce high_________ and one rigid model (i.e., overfitting) model will introduce high _______________.There is a tradeoff between a model's ability to minimize bias and variance.(A) variance, bias (B) bias, bias(C) variance, variance (D) bias, variance12)Which description is not correct in terms of supervised learning, semi-supervisedlearning, unsupervised learning and reinforcement learning?__________(A) Reinforcement learning is one of specific supervised learning methods.(B) Semi-supervised learning falls between unsupervised learning (without anylabeled training data) and supervised learning (with completely labeled training data).(C) Reinforcement learning is neither supervised learning nor unsupervised.(D) Deep reinforcement learning is a combination of deep learning and reinforcementlearning.13)In reinforcement learning, Q-learning defines a function Q(s, a) representing the_______________ when we perform action a in state s, and continue optimally from that point on. The Q-function can be learned iteratively by ______________.(A) The immediate reward plus discounted maximum future reward, Bellman equation(B) Discounted maximum future reward, Markov decision process(C) The immediate reward plus discounted maximum future reward, Markov decisionprocess(D) Discounted maximum future reward, Bellman equation14)When we use one deep convolutional neural network model to classify 101 concepts,which option is not correct in the following description?_________(A)The output of the last fully connected layer can be used as the learning features ofeach concept.(B)The dimension of the classification layer can be 101.(C)The convolutional kernels are pre-defined (i.e., data-independent).(D) Dropout is used to boost the performance.15)Which description is not correct in terms of deep learning?__________(A)Deep learning is essentially a method to learn the features of raw data.(B)Backpropagation is conducted to optimize the weights of deep neural networks sothat the neural network can learn how to correctly map arbitrary inputs to outputs.(C)The achieved performance of deep learning is due to its powerful representationability via many of non-linear mappings.(D)Deep convolutional neural network for classification is employed in an end-to-endmechanism via unsupervised learning.16) Which description is not correct about K-Means clustering?___________(A)K-means clustering can be used for image segmentation and image compression.(B)K is the number of clusters and is generally pre-defined.(C)Each data point can be assigned to more than one cluster.(D)If the dimension of each data points is D, the dimension of cluster centers is D.17)The number of pruned successors in alpha-beta pruning is highly dependent on _____.(A)The moving order(B)The initialized values of alpha and beta(C)The number of terminal nodes(D)Whether breadth-first search or depth-first search is employed18)What description is not correct in terms of AI?(A)Deep learning is one kind of machine learning methods.(B)Machine learning is deep learning.(C)Search is one kind of methods used in AI.(D)In general, LeNet-5 (one of deep convolutional neural networks) maps eachhandwriting images into 0-9 digital character concept space.3.Calculus and Analysis (34 points)1)(Game Playing, 8 points) As shown in the following figure, there is a MINMAXsearch tree with three layers. The utility values for the leaf nodes are respectively displayed at the bottom of the figure.(a) Fill in the blanks for the utility values associated with the tree nodes (i.e., B, C, D) as well as the root node A. (4 points)(b) Draw mark ‘//’ on some branches in the figure to show that they are pruned by the alpha-beta pruning algorithm. (4 points)2) (Boosting, 8 points ) Boosting is a powerful technique for combining multiple “base” classifiers to produce a form of committee whose performance can be significantly better than that of any of the base classifiers. Consider a two-class classification problem, in which the training data comprises 2D input vectors 1,...,N x x along with corresponding binary target variables 1,...,N t t where {1,1}n t ∈+-. Assume that we have trained three base classifiers 12(),()f f x x ,3()f x and the corresponding weighting coefficients 123,,ααα. Please answer:(a) The final classifier learned by boosting can be given by: (4 points)F final (x )=(b) If three base classifiers 12(),()f f x x ,3()f x are shown in the figure and1230.3,0.5,0.7ααα===. Each base classifier partitions the input space into tworegions separated by a linear decision boundary Ωi . The dark regio n with ‘+’ means the target value is +1 and the bright region with ‘-’ means the target value is -1. Three decision boundaries have been put together in the last right figure and separate the space into six sub-regions, please mark the final decision result in the figure with ‘+’ or ‘-’ for each sub -region.(4 points)3) (Image Restoration, 6 points ) Please share several key tricks that effectively improve the performance of your Image Restoration Algorithm in Project 2. (About 100~150 words).4)(Deep learning, 12 points)(a) Convolution is very important in deep convolutional neural network. Please calculate the convolved value of the center pixel in Figure (1) with the given convolutional kernel in Figure (2). (3 points)(b) Given a single depth slice in Figure (3), please give out the average-pooling value of this slice with 2×2 filters and stride 2. (3 points)(c) If we trained a deep convolutional neural network as follows in Figue (4). The sofmax is used to classify five concepts (e.g., car, airplane, truck, ship and person). If we input a car image into the trained deep model, please write out one of likely 5-dimensional outputs by the deep model. (3 points)Figure (4)(d) Please write down the trainable parameters in this model. (3 points)《Artificial Intelligence》Final Examination Answer SheetName: Student ID: Dept.: _1.Fill in the blanks (30 points, 2pt/per)1).______________________, _______________________2). _____________________, _______________________3)._____________________4). _____________________, _______________________5)._____________________6). _____________________7)._____________________, _______________________8)._____________________9)._____________________10).____________________ , _______________________2.Multiple Choice (36 points, 2pt/per)3. Calculus and Analysis (34 points)1) (Game Playing, 8 points)(a) (4 points)(b)(4 points)2) (Boosting, 8 points)(a) (4 points)F final(x)=(b) (4 points)3) (Image Restoration, 6 points)4) (Deep Learning, 12 points)(a)(3 points)(b)(3 points)(c)(3 points)(d)(3 points)。
人工智能的基础知识近年来,人工智能(Artificial Intelligence,简称AI)作为一个热门话题经常被提及。
人工智能是一门研究如何使机器能够像人一样进行智能决策与行动的科学。
在探索人工智能的世界中,存在着一些基础知识,这些知识是理解和掌握人工智能的基石。
首先,人工智能的核心概念之一是机器学习(Machine Learning)。
机器学习是一种通过从数据中学习经验规律,而无需明确编程来让机器具备学习能力的方法。
机器学习基于数学和统计学理论,通过训练算法来自动发现数据中的模式和规律。
这些训练过程使得机器能够从大量数据中获取知识,并在面对新的情况时作出准确的决策。
随着机器学习的发展,许多应用领域如自然语言处理、计算机视觉和语音识别等都取得了重大突破。
其次,人工智能领域的另一个重要概念是深度学习(Deep Learning)。
深度学习是机器学习的一种特殊领域,其灵感来源于人脑神经网络的结构。
深度学习通过建立多层神经网络模型来学习和抽取数据中的特征,并利用这些特征进行预测和决策。
深度学习的算法在图像识别、自然语言处理和语音识别等领域中取得了令人瞩目的成果,从而推动了人工智能的快速发展。
此外,人工智能中还有重要的应用领域,例如自然语言处理(Natural Language Processing,简称NLP)。
NLP关注如何使计算机能够理解和处理自然语言,包括语义分析、实体识别、问答系统等。
通过对大量的语言数据进行分析和建模,NLP 能够使计算机能够进行智能文本理解和生成,这对于聊天机器人、智能助手和智能翻译等应用非常重要。
人工智能也与计算机视觉(Computer Vision)有密切的联系。
计算机视觉旨在使计算机能够理解和解释图像和视频数据。
通过使用图像处理、模式识别和机器学习的技术,计算机视觉可以实现图像分类、物体检测和图像生成等任务。
计算机视觉的应用非常广泛,如自动驾驶、人脸识别和视频监控等。
ai人工智能原理AI人工智能原理人工智能(Artificial Intelligence,AI)是一门致力于研究、设计和实现智能化机器的科学与技术。
AI人工智能的发展源于对人类智能的模拟与复制,旨在使机器能够像人一样具备思维、学习、推理和决策等智能能力。
其核心原理主要包括机器学习、深度学习、自然语言处理和专家系统等。
机器学习是AI人工智能的重要支撑,它是指通过对大量数据的学习和分析,使机器能够从中获取知识和经验,并利用这些知识和经验来解决问题。
机器学习的核心思想是通过建立模型,对输入数据进行分析和预测。
常见的机器学习算法包括决策树、支持向量机和神经网络等。
机器学习的应用广泛,例如在医疗诊断、金融风控和智能推荐等领域都有着重要的作用。
深度学习是机器学习的一个重要分支,它模仿人脑的神经网络结构,构建起多层次的神经网络模型。
深度学习通过多层次的特征提取和抽象,能够自动学习复杂的非线性关系,实现对大规模数据的高效处理和分析。
深度学习在计算机视觉、语音识别和自然语言处理等领域取得了重大突破,推动了AI人工智能的发展。
自然语言处理(Natural Language Processing,NLP)是研究如何使机器能够理解和处理人类自然语言的一门学科。
NLP包括文本分析、语义理解、情感分析和机器翻译等技术,通过对语言的分析和处理,使机器能够与人进行自然的对话和交流。
NLP的应用广泛,例如智能客服、智能助手和智能翻译等领域都离不开NLP的支持。
专家系统是一种基于规则和知识的推理和决策系统。
它通过将专家的知识和经验转化为规则或知识库,并利用规则推理和知识推理的方法,实现对问题的分析和解决。
专家系统在医疗诊断、故障诊断和决策支持等领域发挥着重要作用,能够帮助人们快速准确地做出决策。
AI人工智能的发展离不开大数据和计算能力的支持。
大数据提供了海量的数据资源,为机器学习和深度学习提供了数据基础。
计算能力的提升则使得机器能够更加高效地进行计算和处理,加速了AI人工智能的发展进程。
人工智能概述人工智能(Artificial Intelligence,简称AI)是一门研究如何使计算机具有智能行为的学科,旨在模拟人类智能的思维和行为。
它涉及到多个领域,包括机器学习、自然语言处理、计算机视觉等。
近些年来,人工智能在各行各业得到了广泛应用,如医疗诊断、智能交通、智能家居等。
一、人工智能的背景与发展人工智能的起源可以追溯到上世纪50年代,随着计算机科学的发展,人们开始尝试开发能够模拟人类思维的计算机程序。
随着硬件技术与算法的不断进步,人工智能得到了长足的发展,逐渐具备了一定的自主学习和推理能力。
二、人工智能的基本原理与方法1. 机器学习:机器学习是人工智能的核心技术之一,通过让计算机从大量数据中进行学习和预测,从而使其具备自动识别和分类的能力。
2. 自然语言处理:自然语言处理是指让计算机能够理解和处理自然语言的技术。
它可以用于语音识别、机器翻译、智能客服等领域。
3. 计算机视觉:计算机视觉致力于让计算机能够感知和理解图像和视频内容,从而实现人机交互、图像识别等应用。
三、人工智能的应用领域1. 医疗诊断:人工智能在医疗领域的应用已经取得了显著的突破,能够辅助医生进行疾病诊断、个性化治疗方案制定等。
2. 智能交通:人工智能可以优化交通运输系统,提高路况监测、交通信号控制等效率,减少交通拥堵和事故发生。
3. 智能家居:通过人工智能技术,可以实现家居设备的智能化管理,如语音控制、自动化调控等。
4. 金融领域:人工智能在金融领域的应用非常广泛,可以进行风险评估、投资建议、反欺诈等工作。
四、人工智能的挑战与展望尽管人工智能在许多领域都取得了显著的进展,但仍然存在一些挑战。
例如,数据隐私和安全问题、算法的不透明性、伦理和道德问题等。
未来,人工智能将继续发展并与更多领域相结合,为人们创造更多智能化、便捷化的应用。
总结:人工智能是一门致力于实现计算机智能化的学科,经过多年的发展,已经在各个领域得到了广泛应用。
基于专利信息的人工智能技术创新网络图谱研究赵程程(上海工程技术大学管理学院,上海201620)摘要:本文使用德温特创新专利引文索引数据库提供的人工智能相关专利数据(2010—2020年),利用知识图谱软件Citespace绘制人工智能技术创新网络图谱,识别出AI创新关键路径及重要节点企业/科研机构、中国人工智能领域最有发展潜力的创新主体和重要创新主体,并对其创新合作特征进行分析。
结论如下:①2010—2020年人工智能关键路径及演化呈现四大特征;②AI创新关键路径重要节点企业/科研机构有15家;③AI最具潜力的创新主体有华为等10家企业,以及中山大学等15家高校/科研院所;④AI重要创新主体有腾讯等10家企业,以及电子科技大学等10家科研机构。
关键词:人工智能;创新网路;专利分析;知识图谱中图分类号:G327.312.6文献标识码:AResearch on Artificial Intelligence Technology Innovation NetworkKnowledge Map Based on Patent InformationZhao Chengcheng(School of Management,Shanghai University of Engineering Science,Shanghai201620,China)\'-I r.ici:Firstly,we use the artificial intelligence(AI)patent data(2010一2020)provided by Derwent Innovations Index.Secondly, AI technology innovation network knowledge map was drawn by the software CiteSpace.Thirdly,AI innovation key path and important nodes were identified.Fourthly,the most potential innovation subjects and important innovation subjects in the field of AI in China are analyzed.The conclusions are as follows.①There are4characteristics of evolution in AI key path in2010一2020.②There are15the keynote enterprises/scientific research institutions on AI technology innovation.③The most potential innovation subjects of AI in China include10enterprises such as Huawei,and15scientific institutes such as Sun Yat-Sen University,Xi'an Jiaotong University,Southeast University and Shenzhen University.④20AI innovation sectors in China include10enterprises such as Tencent,and10scientific research institutions such as University of Electronic Science and technology.Key words:Artificial intelligence;Innovation network;Patent analysis;Knowledge map疫情防控的严峻形势正倒逼着各国科技企业人工智能既是机遇又是挑战,世界格局极有可能竞相发力,对错过前三次科技革命的中国来说,因此而重新洗牌。
人工智能详细科普AI(人工智能(Artificial Intelligence))一般指人工智能(计算机科学的一个分支)人工智能(Artificial Intelligence),英文缩写为AI。
它是研究、开发用于模拟、延伸和扩展人的智能的理论、方法、技术及应用系统的一门新的技术科学。
人工智能是计算机科学的一个分支,它企图了解智能的实质,并生产出一种新的能以人类智能相似的方式做出反应的智能机器,该领域的研究包括机器人、语言识别、图像识别、自然语言处理和专家系统等。
人工智能从诞生以来,理论和技术日益成熟,应用领域也不断扩大,可以设想,未来人工智能带来的科技产品,将会是人类智慧的“容器”。
人工智能可以对人的意识、思维的信息过程的模拟。
人工智能不是人的智能,但能像人那样思考、也可能超过人的智能。
人工智能是一门极富挑战性的科学,从事这项工作的人必须懂得计算机知识,心理学和哲学。
人工智能是包括十分广泛的科学,它由不同的领域组成,如机器学习,计算机视觉等等,总的说来,人工智能研究的一个主要目标是使机器能够胜任一些通常需要人类智能才能完成的复杂工作。
但不同的时代、不同的人对这种“复杂工作”的理解是不同的。
定义详解人工智能的定义可以分为两部分,即“人工”和“智能”。
“人工”比较好理解,争议性也不大。
有时我们会要考虑什么是人力所能及制造的,或者人自身的智能程度有没有高到可以创造人工智能的地步,等等。
但总的来说,“人工系统”就是通常意义下的人工系统。
关于什么是“智能”,就问题多多了。
这涉及到其它诸如意识(CONSCIOUSNESS)、自我(SELF)、思维(MIND)(包括无意识的思维(UNCONSCIOUS_MIND))等等问题。
人唯一了解的智能是人本身的智能,这是普遍认同的观点。
但是我们对我们自身智能的理解都非常有限,对构成人的智能的必要元素也了解有限,所以就很难定义什么是“人工”制造的“智能”了。
因此人工智能的研究往往涉及对人的智能本身的研究。
专题:大力推进科研范式变革Vigorously Promote Scientific Research Paradigm Transform引用格式:李国杰. 智能化科研(AI4R):第五科研范式. 中国科学院院刊, 2024, 39(1): 1-9, doi: 10.16418/j.issn.1000-3045.20231007002.Li G J. AI4R: The fifth scientific research paradigm. Bulletin of Chinese Academy of Sciences, 2024, 39(1): 1-9, doi: 10.16418/j.issn.1000-3045.20231007002. (in Chinese)编者按随着大数据与人工智能(AI)技术的飞速发展,人类正迎来新一轮科技革命与产业变革。
深度学习等技术近年来的突破,也使AI在数学、物理学、化学、生物学、材料学、制药等自然科学和高技术领域的研究中得到了广泛应用并取得了令人瞩目的重大成果。
AI的快速发展为人类的科学研究工具和组织模式的效率提升提供了新机遇,以AlphaFold2和ChatGPT为代表的智能工具,展现出了超越人类解决复杂问题的能力。
趋势表明,AI for Science正在成为一种新的科研范式。
智能时代已经到来,科研范式与形态的变革刻不容缓,我们必须把握机遇,积极应对。
为此,《中国科学院院刊》特组织策划专题“大力推进科研范式变革”,本专题由《中国科学院院刊》副主编、中国工程院院士、中国科学院计算技术研究所李国杰研究员指导推进。
智能化科研(AI4R):第五科研范式李国杰中国科学院计算技术研究所北京100190摘要文章将“智能化科研”(AI4R)称为第五科研范式,概括它的一系列特征包括:(1)人工智能(AI)全面融入科学、技术和工程研究,知识自动化,科研全过程的智能化;(2)人机智能融合,机器涌现的智能成为科研的组成部分;(3)有效应对计算复杂性非常高的组合爆炸问题;(4)面向非确定性问题,概率统计模型在科研中发挥更大的作用;(5)跨学科合作成为主流科研方式,实现前4种科研范式的融合;(6)科研更加依靠以大模型为特征的科研大平台等。
浙江大学人工智能研究所(以下简称AI)创建于1981年,是专门从事科学研究与培养高层次计算机专业人才的科研机构。
全所现有各类研究人员46名,其中教授15名(含中国工程院院士1名,长江计划特聘教授1名,博士生导师10名),副教授22名,讲师10名。
同时设有计算机应用技术博士点、硕士点和计算机科学与技术博士后流动站。
现任研究所所长为浙江大学校长、中国工程院院士潘云鹤教授,副所长为董金祥教授、朱淼良教授和周昌乐教授,学术委员会主任为何志均教授。
研究所下设:
知识工程研究室、智能CAD研究室、CAD&CG研究室、计算机视觉与智能机器人研究室、智能信息管理与决策研究室等5个研究室、实验室,以及资料室、办公室等机构。
主要研究方向:
人工智能理论,形象思维,计算机图形学(CG)与计算机辅助设计(CAD),计算机集成制造(CIMS)及其它先进制造技术,智能CAD,信息智能和决策支持,计算机视觉与智能机器人,多媒体技术,工程数据库,智能控制,计算机网络和信息通讯,科学可视化,分布式知识库,操作系统,数据库,管理信息系统,计算机辅助工业设计等。
学术带头人:
Founded in 1981, the Artificial Intelligence Research Institute (AIRI) is an institution specializing in scientific research and high-level computer talent cultivating. Presently it has a staff of 46. Among them, there are 15 professors ( including one member of Chinese Engineering Academy, one member of China Education Ministry’s Cheung Kong (Chiang Jiang) Scholar,ten doctoral supervisors), 22 associate professors and 10 lecturers. In addition to the Ph.D and Master degree programs, it also offers post doctoral program in computer science and technology. Chief of Institution is Prof. Pan Y unhe, who is also president of Zhejiang University and a member of Chinese Engineering Academy. Prof. Dong Jinxiang and Prof. Zhu Miaoliang and Prof. Zhou Changle are vice chiefs of the Institution. Prof. He Zhijun is chairman of academia committee of AIRI.
The AIRI consists of one reference room, one administrative office and five research labs, including Knowledge Engineering Lab, Intelligence CAD Lab, CAD&CG Lab, Computer vision & Robotics Labs, Intelligence Information Management and Decision Lab. The main research scope is as follows:
Artificial Intelligence theory
Imagery thinking
Computer Graphics and Computer Aided Design
Computer Integrated Manufacture system (CIMS) and other Advanced Manufacturing Technologies
Intelligence CAD
Information Intelligence and policy support
Computer Vision and Robotics
Multimedia Technology
Engineering Database
Intelligent Control
Computer Network and Communication
Scientific Visualization
Distributed Knowledge Database
Operate System
DataBase
Management Information System Computer Aided Industry Design Professors:。