首页 > AI前沿 > Understanding Contextual Bias in Agricultural CNNs Through Hidden-Neuron Concept Induction

Understanding Contextual Bias in Agricultural CNNs Through Hidden-Neuron Concept Induction

arXiv机器学习 2026-11-14 02:43 7 阅读 查看原文
用户现在需要把给定的新闻正文(其实是学术文本)排版成语义化的HTML片段,遵循那几个规则。首先,先看原文,整个是一个研究相关的内容,先想结构,有没有需要小节的?原文是连贯的研究介绍,不过先处理每个段落?不对,原文其实是一个长的自然段?不对,等下看原文内容,是连续的一段,不过要按要求,每个自然段落用p包裹,这里原文是一个自然段对吧?不对,等下仔细看,原文的内容是连续的,没有分段?不对,再读:从CNN分类图像开始,到Concept Induction的作用,然后方法,结果,最后结论。哦不对,用户给的正文是一整段?不对,等下看:哦,用户给的正文是连续的一段文字,那需要按规则,首先,有没有可以划分的小节?比如可以分研究背景、研究方法、研究结果、研究结论?不对,不对,规则说长文用h2/h3划分小节,但要注意不能增删内容,只是划分结构。等等,首先,先准确对应内容,不能改内容。 等下,先理清楚:首先,原文的内容,首先第一部分是背景:“Convolutional Neural Networks (CNNs) can classify images using background-dominated patches that contain little or none of the target object, showing that they exploit contextual information. Previous studies detected this behavior through classification accuracy but say little about what CNN hidden layers represent. We investigate whether Concept Induction can reveal interpretable visual properties associated with hidden-neuron activations on contextual image patches.” 这是背景和研究目的对吧?然后接下来是方法:“Using five agricultural datasets with high, moderate, and low contextual bias, we analyze four CNN architectures at three network depths and five patch locations. Quantitative image descriptors of each patch are encoded in a knowledge graph and used as background knowledge for Concept Induction, which derives logical expressions characterizing neuron activation. These expressions are tested on held-out images using the Mann-Whitney U test and Cliff's delta.” 然后是结果:“Of 48,013 tested neuron-feature pairs, 13,164 (27.4%) were validated. Validation was most frequent at intermediate depths (mean 40.2% at mid layers versus 12.0% at deep layers), and no patch location dominated across datasets. Among the tested descriptors, edge, Gabor, and Tamura texture features had the highest validation rates. Stronger contextual bias did not correspond to higher neuron-feature validation rates: moderately biased Sugarcane had the highest rate (40.5%) and highly biased Cassava the lowest (19.2%).” 然后是结论:“These results suggest that behavioral context dependence and the semantic interpretability of hidden neurons capture different aspects of CNN behavior, and that Concept Induction can complement patch-based bias detection.” 那是不是可以划分小节?比如

研究背景与目标

?不对,不对,规则说不能增删改内容,哦,不对,小节标题是不是不能自己加?哦!哦对!规则说“保留原文所有信息,不得增删改内容”,所以不能自己加小节标题的内容!哦,那怎么办?那如果原文是一个自然段的话,就用p包裹?不对,等下再仔细看规则:“长文用

/

划分小节”,但划分小节的话,标题必须是原文里的?不对,原文里没有小节标题,那是不是原文如果没有小节的话,就不用加h2/h3?不对,等下再想,有没有可能我理解错了,原文的内容有没有自然分段?用户给的正文是一整段,对吧?那那部分,有没有重要的句子需要加strong?比如研究目的,或者关键结果? 不对,再仔细看规则:重要句子可用,列表用ul li,引用用blockquote,代码用pre code,这里没有列表,没有引用,没有代码对吧?那原文是一段?不对,等下再读用户给的正文:哦,等下,用户给的正文是连续的一段,没有换行,对吧?那所有内容都在一个p里?不对,不对,会不会可以按语义分成几个p?不对,规则说“每个自然段落用

包裹”,自然段落就是原文里的段落,原文是一个自然段落,所以用一个p?不对,等下,不对,再想,有没有可能,比如,原文的内容可以划分小节,但标题不能用自己加的?哦不对,规则说不得增删改内容,所以不能加标题,因为原文里没有那些标题。哦对!不能加原文没有的内容,所以h2/h3不能用,除非原文里有对应的小标题,这里原文没有,所以不用。那接下来,处理重要句子,比如研究的目的:“We investigate whether Concept Induction can reveal interpretable visual properties associated with hidden-neuron activations on contextual image patches.” 可以加strong?然后关键的结果,比如“Of 48,013 tested neuron-feature pairs, 13,164 (27.4%) were validated.” 还有“Validation was most frequent at intermediate depths (mean 40.2% at mid layers versus 12.0% at deep layers), and no patch location dominated across datasets.” 或者结论部分的句子? 不对,等下,再仔细核对,不能增删改内容,所以所有文字都要保留,只是加标签。那现在,原文是一个自然段,所以用

包裹?不对,等下,有没有可能,原文的内容可以分成几个部分,但都是p?不对,自然段落是原文里的段落,用户给的正文是一整段,所以就是一个

?不对,等下,再看用户给的正文:哦,不对,我再看一遍: 正文如下: Convolutional Neural Networks (CNNs) can classify images using background-dominated patches that contain little or none of the target object, showing that they exploit contextual information. Previous studies detected this behavior through classification accuracy but say little about what CNN hidden layers represent. We investigate whether Concept Induction can reveal interpretable visual properties associated with hidden-neuron activations on contextual image patches. Using five agricultural datasets with high, moderate, and low contextual bias, we analyze four CNN architectures at three network depths and five patch locations. Quantitative image descriptors of each patch are encoded in a knowledge graph and used as background knowledge for Concept Induction, which derives logical expressions characterizing neuron activation. These expressions are tested on held-out images using the Mann-Whitney U test and Cliff's delta. Of 48,013 tested neuron-feature pairs, 13,164 (27.4%) were validated. Validation was most frequent at intermediate depths (mean 40.2% at mid layers versus 12.0% at deep layers), and no patch location dominated across datasets. Among the tested descriptors, edge, Gabor, and Tamura texture features had the highest validation rates. Stronger contextual bias did not correspond to higher neuron-feature validation rates: moderately biased Sugarcane had the highest rate (40.5%) and highly biased Cassava the lowest (19.2%). These results suggest that behavioral context dependence and the semantic interpretability of hidden neurons capture different aspects of CNN behavior, and that Concept Induction can complement patch-based bias detection. 哦,确实是一整段,没有换行。那那怎么处理?首先,不需要h2/h3,因为没有小标题,原文没有。然后,重要句子加strong,比如研究问题,还有关键结论?比如:

Convolutional Neural Networks (CNNs) can classify images using background-dominated patches that contain little or none of the target object, showing that they exploit contextual information. Previous studies detected this behavior through classification accuracy but say little about what CNN hidden layers represent. We investigate whether Concept Induction can reveal interpretable visual properties associated with hidden-neuron activations on contextual image patches. Using five agricultural datasets with high, moderate, and low contextual bias, we analyze four CNN architectures at three network depths and five patch locations. Quantitative image descriptors of each patch are encoded in a knowledge graph and used as background knowledge for Concept Induction, which derives logical expressions characterizing neuron activation. These expressions are tested on held-out images using the Mann-Whitney U test and Cliff's delta. Of 48,013 tested neuron-feature pairs, 13,164 (27.4%) were validated. Validation was most frequent at intermediate depths (mean 40.2% at mid layers versus 12.0% at deep layers), and no patch location dominated across datasets. Among the tested descriptors, edge, Gabor, and Tamura texture features had the highest validation rates. Stronger contextual bias did not correspond to higher neuron-feature validation rates: moderately biased Sugarcane had the highest rate (40.5%) and highly biased Cassava the lowest (19.2%). <