||
We all know that there are no Nobel Prizes for the field of computer science and engineering because this field is still relatively new. Mr. Alfred Nobel gave his will in 1895 contributing to funds for what is now called Nobel Prizes. If we consider Alan Turing’s paper in 1936 as the first mathematical model for general-purpose computers, the field of computer science and engineering was not born till 41 years after the Nobel’s will. Modern Turing Awards, each year carying an amount of about one-million dollars in prize, were widely considered of a Nobel Prize caliber. However, we might not want to regard awards of similar amounts of award money to be comparable in quality.
我们都知道,计算机科学与工程领域没有设诺贝尔奖,因为该领域还是一个相对较新的领域。阿尔弗雷德·诺贝尔(Alfred Nobel)先生于1895年写了自己的遗嘱,设立了当代的诺贝尔奖。如果我们将阿兰·图灵(Alan Turing)1936年的论文视为通用计算机的第一个数学模型,则计算机科学和工程领域的诞生要到诺贝尔遗嘱之后的第41年。每年获得约一百万美元奖金的现代图灵奖被广泛地认为是和诺贝尔奖相当。但是,我们可能不希望将金额数作为比较奖的质量的尺度。
Here is an example with the latest Turing Award 2018. I raise this example publicly here so that we can conduct a fruitful academic discussion. By conducting the discussion, we are exercising our basic rights for freedom of speech as well as academic freedom.
2018年图灵奖就是一个例子。我在这里公开提出这个例子,以便我们进行富有成果的学术讨论。通过进行讨论,我们正在行使言论自由和学术自由的基本权利。
Under William-Webster online dictionary, the definition of plagiarizing is “to steal and pass off (the ideas or words of another) as one’s own: use (another's production) without crediting the source.”
在威廉-韦伯斯特在线词典中,剽窃的定义是“窃取和冒充(他人的想法或话语)作为自己的东西:使用(他人的成果)而不引用其来源”。
Did Turing Awards 2018 Give to Plagiarism? Has the work of Turing Prize 2018 violated the Code of Ethics and Professional Conduct (CEPC) of the granting organization ACM? ACM CEPC code stipulates: “1.5: … credit the creators of ideas, …”
2018年图灵奖给了犯了剽窃的工作吗?2018年图灵奖的工作是否违反了授予组织ACM的《道德与职业行为准则》(CEPC)?ACM CEPC准则规定:“ 1.5:…引用思想的原创者…”
I invite the reader to judge for himself, and provide his comments on my Facebook page, whether the ACM Turing Award 2018 work plagiarized Cresceptron. To spell out the context clearly, let me presents Evidence 1 first.
我请读者自己进行判断,并在我的脸书页面(或科学网也行)发表他的评论,以确定ACM 2018年图灵奖的工作是否剽窃了生长网(Cresceptron)。为了清楚地说明上下文,让我首先介绍证据1。
Evidence 1: State of knowledge before Cresceptron in 1993. Before 1993 many researchers thought it was impossible to recognize 3D objects from 2D images using neural networks without any monolithic 3D object model. 3D-model based aspect graphs were the norm in the computer vision community then for recognizing 3D objects. Cresceptron, published in 1993, completely changed the prevailing perspective, and set the trend of general purpose vision (detection, recognition, and segmentation) from cluttered 3D scenes, which ImageNet and many later projects are based upon.
证据1:1993年Cresceptron之前的知识状态。1993年之前,许多研究人员认为,如果没有任何整体3D物体模型,就无法使用神经网络从2D图像识别3D物体。基于3D模型的方面图是当时计算机视觉领域中用于识别3D对象的广用方法。于1993年发表的Cresceptron彻底改变了这个主流观点,并从杂乱的3D场景中为通用视觉(检测,识别和分割)启动趋势。ImageNet和许多后来项目都用了此类并类杂乱的3D场景。
Evidence 2: Cresceptron. The first published work of recognizing 3D objects from 2D images using deep learning networks is Cresceptron:
证据2:Cresceptron。首次发表的使用深度学习网络从2D图像识别3D对象的工作是Cresceptron:
[0] J. Weng, N. Ahuja and T. S. Huang, “Cresceptron: a self-organizing neural network which grows adaptively,” in Proc. Int'l Joint Conference on Neural Networks, Baltimore, Maryland, vol. 1, pp. 576-581, June 1992.
[1] J. Weng, N. Ahuja and T. S. Huang, “Learning recognition and segmentation of 3-D objects from 2-D images,” in Proc. 4th International Conf. Computer Vision (ICCV), Berlin, Germany, pp. 121-128, May 1993.
[2] J. Weng, N. Ahuja and T. S. Huang, “Learning recognition and segmentation using the Cresceptron,” International Journal of Computer Vision (IJCV), vol. 25, no. 2, pp. 105-139, Nov. 1997.
For an intuitive view about the trend-setting idea of learning 3D vision from 2D images without a 3D object model, see Figure 2 in [2].
关于在没有3D物体模型的情况下为从2D图像学习3D视觉的启动趋势的思想的直观视图,请参见[2]中的图2。
As far as I know,the Cresceptron is the first in the following categories, from most important (1) to the least important (6):
(1) The First deep learning network for 3D objects from 2D images using learned 2D features without a 3D model.
(2) The First deep learning network that performs both detection and recognition, using regular scans with strides and using different scales.
(3) The First deep learning network that avoids matching large-scale 2D patterns by reasoning matching large-scale 2D patterns is unlikely.
(4) The first deep learning network that proposed a pairing architecture: a convolution layer paired with a resolution-reduction (subsampling) layer.
(5) The first deep learning network that originated tolerance ideas in the pairing architecture: what is now called max-pooling (maximization) idea and blurring idea (weighted average) for the resolution-reduction layers.
(6) The First deep learning network that conducts segmentation, using a post-analysis of contributing features.
据我所知,Cresceptron在以下几类思想的原创,从最重要的(1)到最不重要的(6):
(1)第一个从2D图像中学习3D物体的深度学习网络,而没有一个3D模型。用的方法是学来的2D特征。
(2)第一个同时进行检测和识别的深度学习网络。用的方法是跨步定格扫描并使用不同的比例。
(3)第一个避免匹配大尺寸的2D模板的深度学习网络。理由是大尺寸的2D模板匹配好不太可能。
(4)第一个配对架构的深度学习网络:卷积层与分辨率降低(子采样)层配对。
(5)第一个原创了配对架构中容差思想的深度学习网络:现在称为分辨率降低层中的最大汇集(max-pooling)思想和模糊思想(用加权平均)。
(6)第一个做分割的深度学习网络。使用的方法是贡献特征进行后分析。
It is unlikely the awardees were not aware of the above two papers since ICCV and IJCV were considered the best-known computer vision conference and the best computer vision journal, respectively, by many researchers.
由于ICCV会议和IJCV期刊被许多研究员分别认为是最著名的计算机视觉会议和最佳计算机视觉杂志,因此获奖者不太可能没有读上述两篇论文。
Evidence 3: In his slide titled “Further discussions of it merely encumbers the literature,” Hinton’s accounts of the rejections of their neural network papers in 2007, 2009 and 2010 which were 14 years, 16 years, and 17 years, respectively, after Cresceptron was published in ICCV. According to their publication records, Hinton did not do 3D vision work until 2001, 8 years after Cresceptron. Before that, Hinton was interested in physics-based models, contrary to ideas (1) through (6). LeCun did not publish 3D vision work until 2004, 11 years after Cresceptron. Before that LeCun was focusing on 2D document processing, primarily individual character recognition. As a postdoc of LeCun, Yoshua Bengio was also doing 2D document processing, mainly individual character recognition.
证据3:在题为“对它的进一步讨论仅是文章受阻”的幻灯片里Hinton讨论了2007年、2009年和2010年,即Cresceptron在ICCV发表之后的14年,16年和17年,其几篇神经网络论文遭拒。根据他们的出版物记录,Hinton直到2001年,即Cresceptron的8年后才进行3D视觉工作。在此之前,Hinton对基于物理的模型很感兴趣,这与以上思想(1)至(6)相反。LeCun直到2004年,即Cresceptron发表11年后,才发表3D视觉工作。在此之前,LeCun专注于2D文档处理,主要是单个字符识别。作为LeCun的博士后,Yoshua Bengio当时也进行2D文档处理,主要也是单个字符识别。
However, when doing individual 2D character recognition using CNN (Convolution Neural Network), the Respondents were much later than Kunihiko Fukushima whose 2D-from-2D CNN called Cognitron, although it did not learn at all, was published in Biological Cybernetics in 1975. Paul Werbos first used error backpropagation for neural networks in 1981, noticably earlier than the three Respondents.
但是,当使用CNN(卷积神经网络)进行单个2D字符识别时,被告人要比福岛邦彦(Kunihiko Fukushima)晚得多。福岛邦彦的称为Cognitron的从2D到2D 的CNN根本不做机器学习,于1975年在Biological Cybernetics期刊上发表。Paul Werbos于1981年首次将误差反向传播用于神经网络,这明显早于三位被告人。
Evidence 4: Fukushima plus Werbos did not set a trend. Without Cresceptron’s revolutionary and trend-setting ideas (1) through (6) above, Fukushima’s 2D-from-2D CNN architecture plus Werbos’ error backpropagation could not set such a trend.
证据4:比福岛邦彦和Werbos的工作没有形成趋势。 如果没有上述Cresceptron的革命性和引领趋势的思想(1)至(6),福岛邦彦的2D至2D CNN架构以及Werbos的误差反向传播无法形成现今这种趋势。
Fukushima’s Neocognitron did not do (1) through (6). In particular, Fukushima’s Neocognitron did not use the paired architecture (4) and did not do max-pooling or weighted average ideas in (5).
特别是,福岛邦彦的Neocognitron并未执行(1)至(6)。 特别是,福岛的Neocognitron没有使用配对架构(4),也没有用(5)中的使用最大池或加权平均思想。
The next episode will discuss large-scale data falsifications with error backpropagation.
下一集将讨论用误差反向传播进行大规模的数据造假。
Evidence 5: Cresceptron’s original ideas (1) through (6) are critical in later published 3D vision work by the three awardees, G. Hinton, Y. LeCun, and G. Bengio. The three Awardees later published 3D vision work in which Cresceptron’s original ideas (1) through (6) not only greatly directed the Awardees’ work but also gave the Awardees confidence to follow this approach. In other words, the three Awardees are primarily imitatorsof Cresceptron; however, the Awardees (e.g., Evidence [3] and [4] below) did not cite Cresceptron in any way, let alone properly credit Cresceptron, as required by CEPC 1.5.
证据5:在三位获奖者Hinton,LeCun和Bengio后来发表的3D视觉工作中,Cresceptron的原创思想(1)至(6)至关重要。随后,三位获奖者发表了3D视觉工作中Cresceptron的原创思想(1)至(6)不仅大大地指导了获奖者的工作,而且使获奖者有信心采用这种方法。换句话说,三个获奖者主要是Cresceptron的模仿者;但是,获奖者(例如下面的证据[3]和[4])没有以任何方式引用Cresceptron,更不用说按照CEPC 1.5的要求那样对Cresceptron给予适当的功劳评价了。
Plagiarism, as defined by the official document titled University of Toronto Code of Behavior on Academic Matters: “the wrongful appropriation and purloining, and publication as one’s own, of the ideas, or the expression of the ideas ... of another.”
根据题为《多伦多大学学术事务行为守则》的正式文件的定义,剽窃是:“对他人思想或对他人思想的表达的不正当挪用和偷窃,作为自己的发表。”
Hinton might argue that his group was only using ImageNet data of 3D scenes. However, ImageNet data did not require them to follow Cresceptron’s main original ideas (1) through (6).
Hinton可能会争辩说,他的团队仅使用ImageNet 的3D场景的数据。但是,ImageNet数据并不要求他们遵循Cresceptron的主要原创思想(1)至(6)。
Evidence 6: Krizhevsky, Sutskever & Hinton 2012 [3] below, relied upon Cresceptron’s original ideas (1) through (6), and even the CNN architecture (See Figure 2 in [3]) follows that of Cresceptron. However, Cresceptron is never cited by [3], let alone properly credited by [3], as required by CEPC 1.5.
证据6:以下Krizhevsky,Sutskever和Hinton 2012 [3]依靠Cresceptron的原创思想(1)至(6),甚至其CNN架构(请参见[3]中的图2)也都遵循了Cresceptron的架构。但是[3]从未引用过Cresceptron,更不用说按照CEPC 1.5的要求恰当地引用Cresceptron了。
[3] A. Krizhevsky, I. Sutskever and G. Hinton, ImageNet Classification with Deep Convolutional Neural Networks, Advances in Neural Information Processing 25, pp. 1106-1114, 2012, MIT Press, Cambridge, MA.
Evidence 7: LeCun, Bengio & Hinton 2015 [4] below grossly violated CEPC 1.5 in a similar way. See, e.g., Fig. 2 in [4] that heavily uses Cresceptron ideas (1) to (6).
证据7:下面的LeCun,Bengio和Hinton 2015 [4]以类似的方式严重违反了CEPC 1.5。参见,例如,[4]中的图2大量使用了Cresceptron的思想(1)至(6)。
[4] Y. LeCun and L. Bengio and G. Hinton, Deep Learning, Nature, vol. 521, pp. 436-444, 2015.
See line 6 of abstract: “Deep convolutional nets”. Line 5, first paragraph, left column, page 436: “Machine-learning systems are used to identify objects in images”. Those are almost exclusively 3D objects. See line 3, right column, first paragraph, page 436: “beating records in image recognition (1-4)” where the earliest cited work 1 (2012) was co-authored by Hinton. But in fact, the first one that created records and then beating the records was Cresceptron (1992 and 1997). See Line 1, last paragraph, left column, page 438, “The architecture of a typical ConvNet (Fig. 2) is structured as a series of stages.” This is essentially an imitation of the first 2D convolution architecture for 3D objects that Cresceptron first published. See the top figure (Figure 2) of page 438: “Inside a convolution network”. It used the original ideas of Cresceptron:
(i) Used Cresceptron’s idea (1) (e.g., the Samoyed cat) from 2D images;
(ii) Used Cresceptron’s detection idea (3) (e.g. see the first panel of Fig. 3);
(iii) Used Cresceptron’s idea (3) avoid matching large patterns directly (see Fig. 3);
(iv) Used Cresceptron’s idea (4) (see multiple paired architecture in Fig. 2);
(v) Used Cresceptron’s idea (5) (see green text max-pooling in Fig. 2);
(vi) Used Cresceptron idea (6) to segment 3D objects in 2D images (Cresceptron used polygon, this paper used a simpler rectangular box, see line 19, left column, page 439, “detecting pedestrians” as Cresceptron did).
参见摘要第6行:“深层卷积网络”。第5行,第一段,左栏,第436页:“机器学习系统用于识别图像中的对象”。这些几乎都是3D物体。参见第3行,右栏,第一段,第436页:“图像识别中的打破记录(1-4)”,其中最早引用的工作(2012)是由Hinton作为合作作者的。但实际上,第一个创建记录然后击败记录的系统是Cresceptron(1992和1997年)。参见第1行,最后一段,左栏,第438页,“典型的ConvNet的体系结构(图2)由一系列阶段构成。”这本质上是对Cresceptron首次发布的为处理3D物体的第一个2D卷积架构的模仿。请参见第438页的顶部图(图2):“在卷积网络内部”。它使用了Cresceptron的原创思想:
(i)使用了 Cresceptron的处理2D图像的思想(1)(例如萨摩耶猫);
(ii)使用了Cresceptron的检测思想(3)(例如,参见图3的第一个面板);
(iii)使用了Cresceptron的思想(3)避免直接匹配大型模版(见图3);
(iv)使用了Cresceptron的思想(4)(参见图2中的多对架构);
(v)使用了Cresceptron的思想(5)(请参见图2中的绿色文本最大汇集);
(vi)使用了Cresceptron的思想(6)在2D图像中分割3D对象(Cresceptron使用了多边形,而本文使用了一个更简单的矩形框,请参见第19行第439页左栏,第439页的“检测行人”,就像Cresceptron所做的那样)。
Evidence 8:Respondents might argue that [4] is a review paper, but LeCun and Bengio appear to have had plagiarized Cresceptron as early as 1998 in their own work:
证据8:被告可能会争辩说[4]是一篇综述文章,但是LeCun和Gengio似乎早在1998年就曾在自己的工作中窃Cresceptron:
[5] Y. LeCun, L. Bottou, and Y. Bengio, and P. Haffner, Gradient-Based Learning Applied to Document Recognition, Proc. of the IEEE, vol. 86, no. 11, Nov. 1998.
For example, see e.g., Fig. 2 in [5] which imitates the original architecture of Cresceptron, especially the pairing architecture that is not in Fukushima then. Page 2284 left column, 2ndparagraph, line 19 reads: “a subsampling layer. … The receptive field of each unit is a 2x2 area in the previous layer’s corresponding feature map. Each unit computes the average of its four inputs …” These are the Cresceptron’s original ideas (3) through (5).
例如,参见[5]中的图2,其模仿了Cresceptron的原创架构,尤其是福岛邦彦不用的配对架构。第2284页左栏第二段第19行显示:“子采样层。…每个单元的感受区域是上一层对应的特征图中的2x2区域。每个单元计算其四个输入的平均值……”这是Cresceptron的原创思想(3)至(5)。
For example, all of their earlier publications did not have the pairing architecture nor the subsampling layer.
例如,他们所有在此以前的发表文章都没有配对架构也没有子采样层。
[6] LeCun et al., Handwritten Digit Recognition with a Back-Propagation Network in NIPS 1989
did not have these ideas (3) through (5) (see Fig. 4 in [6]). Likewise,
没有这些思想(3)至(5)(参见[6]中的图4)。同样地,
[7] LeCun et al. Backpropagation Applied to Handwritten Zip Code Recognition, Neural Computation, 1 pp. 541-551, 1989
did not have these ideas (3) through (5) (see Fig. 2 with caption “Neural architecture” but the caption-corresponding figure was inadvertently exchanged with Fig. 3).
没有这些思想(3)到(5)(请参见带有标题“神经体系结构”的图2,但与标题相对应的图与图错误地交换了)。
[8] Kaoru Ota, Minh Son Dao, Vasileios Mezaris, and Francesco G. B. De Natale. 2017. Deep Learning for Mobile Multimedia: A Survey. ACM Trans. Multimedia Comput. Commun. Appl. 13, 3s, Article 34 (June 2017), 22 pages. DOI: https://doi.org/10.1145/3092831 .
confusingly stated “In 1989, LeCun applied the BP to convolutional NN
with adaptive connections, one of the key elements of modern deep learners [65]. In
its architecture, LeCun organized the network in two types of layers, convolutional
and subsampling, each one showing a topographic structure. A couple of years later,
Hochreiter was the first to identify the fundamental DL problem, also known as the
“long time lag problem” [48], which motivated the following works of Bengio et al.
toward a viable solution for deep network learning [21, 49]. More or less in the same
years, the Cresceptron model introduced the use of max-pooling (MP) layers in the
neural architecture [101], a concept that will be later widely adopted in modern DL.”
令人困惑地说:“1989年,LeCun将BP应用于卷积神经网络。其
自适应连接是现代深度学习的关键要素之一[65]。LeCun把其
的体系结构分为两类,即卷积层和子采样层,每个显示一个拓扑结构。几年后
Hochreiter是第一个识别基本DL问题的人,也被称为“长期滞后问题”
[48],激发了Bengio等人的以下工作,以寻求一种可行的深度网络学习解决方案[21,49]。或多或少在同些年,Cresceptron模型在神经体系结构的设计中引入了最大汇集(MP)层的使用[101] 。这一概念在以后的现代DL中被广泛采用。”
Why confusing? It is chronologically confusing for Ota et al. [8] to state “In its architecture, LeCun organized the network in two types of layers, convolutional and subsampling, each one showing a topographic structure.” This chronologically confusing sentence was only possible correct to mean LeCun’s much later publications (e.g., [5] in 1998) which used the pairing architecture of Cresceptron 1992 without citing Cresceptron. But in [8] this confusing sentence was right after a sentence that ended with citing the 9 years earlier 1989 [7] but [7] did not present any pairing architecture nor subsampling.
为什么令人困惑?Ota等人把按时间顺序搞错了。[8]指出“LeCun把其
的体系结构分为两类,即卷积层和子采样层,每个显示一个拓扑结构。”这个句子只可能指LeCun在后来的出版物(例如1998年的[5]),其中使用了Cresceptron 1992的配对结构而没有引用Cresceptron。但是在[8]中,这个令人困惑的句子是在引用了1989年之前的9年[7]的句子之后,但是[7]没有提供任何配对架构或子采样。
In summary, the pairing architecture was originated from Cresceptron in 1992, but Y. LeCun, L. Bottou, and Y. Bengio plagiarized the pairing architecture as well as the weight average idea of Cresceptron six years later in 1998 [7], more specifically the ideas (3) through (5).
总而言之,配对架构起源于1992年的Cresceptron,但是LeCun,Bottou和Bengio在六年后的1998年[7]剽窃了Cresceptron的配对架构以及Cresceptron的加权平均思想,更明确地讲即是思想(3)到(5)。
Cresceptron dealt with much challenging 3D problems but isolated 2D character recognition by LeCun, Bengio and others in [5] 1998 were much simpler special cases of Cresceptron: 2D and isolated single character in instead of 3D world and cluttered world that Cresceptron dealt with.
Cresceptron处理了许多更具挑战性的3D问题,但LeCun,Bengio和其他人在1998年[5] 进行的单个2D字符的识别是比Cresceptron简单得多的特殊情况:2D和单个的字符代替了Cresceptron处理的3D世界和混杂的世界。
Evidence 9: The Turing Award’s Official Announcement [9] below calls the three Respondents “Fathers of the Deep Learning Revolution”:
证据9:以下图灵奖的官方公告[9]将三名被告称为“深度学习革命之父”:
[9] ACM, “Fathers of the Deep Learning Revolution Receive ACM A. M. Turing Award”, Bengio, Hinton, and LeCun Ushered in Major Breakthroughs in Artificial Intelligence, See ACM official web, https://amturing.acm.org, 2019.
Who are the true “fathers”? The three Respondents who imitated Cresceptron?
谁是真正的“父亲”?模仿Cresceptron的三名被告?
Evidence 10: The official “Selection Criteria” in the “Call for Nomination” Page of the Turing Award states: “Although the long-term influences of the nominee’s work are taken into consideration, there should be a particular outstanding and trend-setting technical achievement that constitutes the principal claim to the award.” The three Respondents apparently do not satisfy these Criteria due to the plagiarism claims stated above.
证据10:图灵奖“征集提名”页面上的官方“甄选标准”指出:“尽管考虑被提名者作品的长期影响,但应该有一种特别出色且形成趋势的技术才构成该奖项主要主张的成就。”由于上述的剽窃报告,三位被告者显然不满足这些条件。
A question the Claimant respectfully submits to the ACM Ethics Committee for investigation and judgment is as follows:
原告谨向ACM道德委员会提出的要求调查和判断的问题如下:
Which work has done “trend-setting” for deep learning? The three Claimants’ work or Cresceptron?
哪项工作为深度学习做了“形成趋势”?这三个被告人的工作还是Cresceptron?
With as much as $1M at its disposal each year, has the Turing Award Committee consulted with major researchers in the neural network community, such as the Nobel Prize committees all do? If yes, what are the records from these major researchers in the neural network community?
图灵奖委员会每年可动用的资金高达100万美元。 它是否像诺贝尔奖委员会那样与神经网络领域的主要研究人员进行过咨询?如果是,那些神经网络领域中主要研究人员的给出的记录是什么?
Evidence 11: A. Karpathy*, G. Toderici*, S. Shetty*, T. Leung*, R. Sukthankar*, and L. Fei-Fei, Large-scale Video Classification with Convolutional Neural Networks, in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 24 – 27, Columbus, Ohio, 2014. (* Google staffers).
证据11:A. Karpathy*, G. Toderici*, S. Shetty*, T. Leung*, R. Sukthankar*, and L. Fei-Fei, Large-scale Video Classification with Convolutional Neural Networks, in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 24 – 27, Columbus, Ohio, 2014. (* 为谷歌员工)。
For example, Fig. 2 used Cresceptron’s
(i) idea (1) (e.g., 3D human bodies in Fig. 2) from 2D images;
(ii) idea (2) (e.g., scan with strides) from 2D images (see Section 3.1. 2ndpara. line 10.)
(iii) idea (3) avoid matching large patterns directly (e.g. see the red layers in Fig. 2);
(iv) idea (4) a pairing architecture(see multiple paired architecture in Fig. 2.);
(v) idea (5) (see blue layers in Fig. 2).
例如,图2使用了Cresceptron的
(i)用2D图像的思想(1)(例如,图2中的3D人体);
(ii)用2D图像的思想(2)(例如,跨步定格 )(请参见第3.1节,第2段第10行)。
(iii)思想(3)避免直接匹配大模板(例如,参见图2中的红色层);
(iv)思想(4)配对架构(请参见图2中的多个配对架构);
(v)思想(5)(请参见图2中的蓝色层)。
Questions and answers:
问题与解答:
Q1:Do you have any evidence that the accused have actually seen the original papers about Cresceptron that you cited in your complaint?
问题1:您是否有任何证据表明被告实际上已经看到了在你投诉中引用的有关Cresceptron的原始论文?
A1:I do not have direct evidence, as an individual claimant I have only very limited resource. A criminal case charged by a government agency to a court needs to prove guilty “beyond reasonable doubt” because the government has the needed resource. A case brought up by an individual to a court is a civil case whose standard of proof is “preponderance of evidence”. I assume that ACM uses the “preponderance of evidence” standard.
答1:我没有直接的证据,作为个人,我的资源非常有限。政府机构向法院提起的刑事诉讼需要“没有合理怀疑地证明”有罪,因为政府拥有所需的资源。个人提起的案件是民事案件,其举证标准是“证据优势”。我假设ACM使用“优势证据”标准。
I would like to provide circumstantial evidence to ACM for “preponderance of evidence”.
我想为ACM提供“优势证据”的间接证据。
Hinton should know me through Cresceptron as he has long been a neural network researcher and Cresceptron is the first deep learning network for 3D visual scenes.
Hinton应该通过Cresceptron认识我,因为他一直是神经网络研究人员,而Cresceptron是第一个用于3D视觉场景的深度学习网络。
Soon after we published Cresceptron at ICCV in 1993, Tomaso Poggio invited me to give a talk about Cresceptron at the Center for Biological and Computational Learning, Massachusetts Institute of Technology, Cambridge, Massachusetts, Oct. 19, 1994.
在1993年我们在ICCV上表Cresceptron之后不久,Tomaso Poggio邀请我在1994年10月19日在麻省理工学院生物与计算学习中心进行关于Cresceptron的演讲。
Poggio invited me to contribute a chapter about Cresceptron in his co-edited book: J. Weng, “Cresceptron and SHOSLIF: Toward comprehensive visual learning,''
in S. K. Nayar and T. Poggio (eds.), Early Visual Learning, Oxford University Press, New York, pp. 183-214, 1996.
Poggio邀请我在他的合作编辑的书中撰写有关Cresceptron的章节:J. Weng,“ Cresceptron和SHOSLIF:迈向全面的视觉学习”,
S. K. Nayar和T. Poggio(编),《早期视觉学习》,牛津大学出版社,纽约,第183-214页,1996年。
It seems unlikely that Hinton, as a researcher in the same area, is drastically more careless than Poggio by not seeing Cresceptron at all. Tracking major trend of his research area is Hinton’s responsibility.
作为同一领域的研究人员,Hinton似乎不可能比Poggio更加不仔细而根本没有读到Cresceptron,追踪他研究领域的主要趋势是Hinton的职责。
Q2:Do you know if the accused have ever cited your work, or cited something that describes your work in-depth that would indicate they were aware of it? If so, please provide citations or copies for us to evaluate.
问题2:您是否知道被告是否曾引用您的作品,或引用过一些深入描述您文章的内容以表明他们已经知道了?如果是这样,请提供引文或副本供我们评估。
A2: No. Let me provide some relevant information.
回答2:不。让我提供一些相关信息。
Hinton directly communicated with me as early as Oct. 22, 2001 and ACM has this evidence.
Hinton早在2001年10月22日就直接与我联系,ACM就有此证据。
Although Poggio demonstrated that he was so interested in Cresceptron that he spent travel money (round-trip air-ticket) and accommodation money (hotel) to invite me to give a talk at MIT. Through my talk, I assume that Poggio and his students could learn from Cresceptron as soon as they can. However, I failed to find any publications in Poggio’s work that cited Cresceptron. See “Deep Learning is Hot: Max-Pooling Plagiarism?”, Brain-Mind Magazine, vol. 3, no. 1, 2014 for more context of Poggio’s behavior which seems to be deliberate.
尽管Poggio证明了他对Cresceptron如此感兴趣,以至于他花了旅费(往返机票)和住宿费(酒店)邀请我到MIT演讲。通过我的演讲,我认为Poggio和他的学生可以尽快学习Cresceptron。但是,我在Poggio的著作中找不到引用Cresceptron的出版物。参见“深度学习热:最大汇集剽窃?”,《脑心智杂志》,第1卷。3号2014年1月1日,以了解有关Poggio行为的更多背景信息。这似乎是故意的。
I used “Hinton and Cresceptron” to search under Google Scholar, I found many papers that cited both Hinton and Cresceptron; many are not authored by me. But they are not authored by Hinton.
我在Google学术搜索下使用“Hinton and Cresceptron”进行搜索,发现很多论文都引用了Hinton和Cresceptron。许多不是我写的。但是它们不是Hinton写的。
I used “LeCun and Cresceptron” as key words, I saw a similar situation.
我用“LeCun and Cresceptron”作为关键词,我看到了类似的情况。
Furthermore, I used “Bengio and Cresceptron”, I also saw a similar situation.
此外,我使用了“ Bengio and Cresceptron”,也看到了类似的情况。
Hinton, LeCun and Bengio seem to have been learning from Poggio’s behavior in their attitude toward Cresceptron.
Hinton,LeCun和Bengio似乎向Poggio学习对Cresceptron的态度。
Q3:Please provide specific claims as to which elements of the accused's publications you allege to be plagiarized from your cited prior work. Please explain how these elements differ from the state of the art as described in the literature cited by the accused. Be sure to also highlight the corresponding elements of your prior work that you specifically allege were plagiarized.
问题3:请提供具体主张,说明您指控被告的文章剽窃了你文章中的哪些元素。请解释这些元素与被告引用的文献中描述的当时技术有何不同。确保还突出显示您特别声称被剽窃了的您先前工作中的相应元素。
A3: To answer, I provide additional detail to my above Evidence 9: See [3] .
回答3:我为上面的证据9中提供了更多详细信息:请参阅[3]。
Page 2, line 2 in Section 2, “The images were collected from the web and labeled by human labelers”. Those are almost exclusively images of 3D objects.
第2页,第2行,第2节,“图像是从网上收集并由人工标记者标记的”。这些几乎完全是3D物体的图像。
Page 2, line 2 in Section 3, “The architecture of our network is summarized in Figure 2. It contains eight learned layers — five convolutional and three fully-connected.” In Figure 2 and its caption on Page 5: see “An illustration of the architecture of our CNN”. Figure 2 used the four original ideas of Cresceptron: (A) Used Cresceptron’s Deep learning for 3D objects from 2D images; (B) Used Cresceptron’s idea of detecting 3D objects in images; (C) Used Cresceptron’s idea to segment 3D objects in 2D images (Cresceptron used polygon, this paper used a simpler rectangular box); (D) Use Cresceptron’s pairing architecture; (E) Used Cresceptron’s max-pooling (see line 6 of abstract” “max-pooling layers”).
第2页,第2行,第3节,“我们的网络的体系结构在图2中进行了概述。它包含八个学习层---五个卷积层和三个完全连接层。” 在图2及其第5页的标题中:请参见“我们的CNN架构说明”。图2使用了Cresceptron的6个原创思想:(1)使用Cresceptron的深度学习技术,从2D图像中学习3D物体。(2)使用Cresceptron的原创思想检测图像中的3D物体;用的方法是跨步定格扫描并使用不同的比例。(3)使用Cresceptron的原创思想,避免匹配大尺寸的2D模板。(4)使用Cresceptron的原创思想,配对架构。(5)使用Cresceptron的最大汇集(请参见摘要的第6行“最大汇集”)。(6) 的深度学习网络在2D图像中分割3D对象(Cresceptron使用多边形,本文使用更简单的矩形框)。
An authority wrote “the alleged evidence Dr. Weng presented in support of all
his major accusations would not be considered meaningful by an informed but neutral,
reasonable individual: it is, at most, circumstantial and some is purely conjecture or coincidence.”
一个当局写了“翁博士提出的所谓证据,从一个知情但中立的人看,不会认为他的主要指控有意义:最多只是偶然的情况,有些纯粹是猜测或巧合。”
Circumstantial evidence is sufficient for “preponderance of evidence”.
间接证据足以证明“证据优势”。
The authority went on: “The severity of these accusations might be sufficient to
be actionable in a court of law as cases of libel and/or defamation against Dr. Weng in both the United States and Canada. … “There was no verbatim or near-verbatim copying, nor did we detect any paraphrasing. …The text is correct that the accused did not cite Cresceptron, but that is because, as we indicated above, they were almost certainly not aware of it. … A query directed to the three accused individuals, asking if they recognized the name “Creseptron” or recalled seeing the prior work resulted in strong denials.”
这个当局继续写道:“这些指控的严重性可能足以在美国和加拿大针对翁博士的诽谤和/或诽谤案件在法院提起诉讼。……“没有逐字或近乎逐字的复制,我们也没有发现直接意译。…文本是正确的,即被告没有引用Cresceptron,但这是因为,正如我们上面指出的,他们几乎肯定没有意识到这一点。…针对三名被告有过询问,询问他们是否认出“Creseptron”这个名字,或者回忆起看到先前的工作,他们都强烈否认了。”
Note “Creseptron” was not spelled correctly as Cresceptron. Why was such an important matter wrong? To hint the awardees to deny? Why was this error occurred in a critical step, but all other instances of Cresceptron were correctly spelled in the authority report?
注意:“Creseptron”的拼写没有按Cresceptron正确拼写。为什么这么重要的事情错了?暗示获奖者否认?为什么在关键步骤中发生此错误,但是在当局报告中其它地方都正确拼写了Cresceptron?
Next, the authority wrote incorrectly: “In 1989, LeCun applied the BP to convolutional NN with adaptive connections, …. [65]. In its architecture, LeCun organized the network in two types of layers, convolutional and subsampling” However, in my Evidence 8 already, I already verified “LeCun et al., Handwritten Digit Recognition with a Back-Propagation Network in NIPS 1989 did not have this pairing architecture.”
接下来,这个当局错误地写道:“在1989年,LeCun将BP应用于具有自适应连接的卷积NN,……[65]。LeCun在其体系结构中将网络分为卷积和子采样两种类型。但是,在我的证据8中,我已经验证了“LeCun等人,NIPS 1989中用反向传播网络的手写数字识别没有这配对架构。”
Then the same authority was ready to claim: “This lack of citation does not support an allegation of plagiarism.… we areunable to determine any motive for them to have elided references to this one particular workduring their years of research—individually or collectively.”
然后,这个权威准备宣称:“这种缺乏引用并不支持对剽窃的指控。……我们无法确定他们在多年研究中(无论是个人还是集体)都没有提及这一特定工作的动机。”
Let us review the above definition of plagiarism: “to steal and pass off … the ideas ... of another … as one’s own … without crediting the source.” The awardees unwillingness of admitting they knew Cresceptron and telling their motives of not admitting does prevent a jury from making a verdict based on the principle of preponderance of evidence in a civil court.
让我们回顾一下剽窃的上述定义:“窃取和冒充他人的想法)……作为自己的东西……不引用其来源。”获奖者不愿承认自己知道Cresceptron和不承认自己的动机,并不能阻止陪审团根据民事证据优先原则作出裁决。
To sue for defamation, one must prove that is all of the following is true: published, false, injurious and unprivileged. By privileged, the U.S. lawmakers have decided that in these (report to an authority) and other situations, which are considered “privileged,” free speech is so important that the speakers should not be constrained by worries that they will be sued for defamation.
要起诉诽谤,必须证明以下所有事实都是真实的:发表,虚假,伤害性和无特权的。美国立法者已经决定,在这里(向当局报告)和其他情况下,被认为是“特权”,言论自由是如此重要,以免说话者不必担心受到诽谤的起诉。
In the next episode, we will discuss an even more severe issue: “Did Turing Awards go to fraud?”
在下一集中,我们将讨论一个更为严重的问题:“图灵奖是否给了造假?”
Archiver|手机版|科学网 ( 京ICP备07017567号-12 )
GMT+8, 2024-9-1 08:18
Powered by ScienceNet.cn
Copyright © 2007- 中国科学报社