多负例对比机制下的跨模态表示学习

为了有效地获取到更有区别性的跨模态表示,提出了一种基于多负例对比机制的跨模态表示学习方法——监督对比的跨模态表示学习（supervised contrastive cross-modal representation learning,SCCMRL）,并将其应用于视觉模态和听觉模态上。SCCMRL分别通过视觉编码器和音频编码器提取得到视听觉特征,利用监督对比损失让样本数据与其多个负例进行对比,使得相同类别的视听觉特征距离更近,不同类别的视听觉特征距离更远。此外,该方法还引入了中心损失和标签损失来进一步保证跨模态表示间的模态一致性和语义区分性。为了验证SCCMRL方法的有效性,基于SCCMRL方法构建了相应的跨模态检索系统,并结合Sub_URMP和XmediaNet数据集进行了跨模态检索实验。实验结果表明,SCCMRL方法相较于当前常用的跨模态检索方法取得了更高的mAP值,同时验证了多负例对比机制下的跨模态表示学习具有可行性。     

Graded semantic and phonological similarity effects in priming: Evidence for a distributed connectionist approach to morphology.

A considerable body of empirical and theoretical research suggests that morphological structure governs the representation of words in memory and that many words are decomposed into morphological components in processing. The authors investigated an alternative approach in which morphology arises from the interaction of semantic and phonological codes. A series of cross-modal lexical decision experiments shows that the magnitude of priming reflects the degree of semantic and phonological overlap between words. Crucially, moderately similar items produce intermediate facilitation (e.g., lately-late). This pattern is observed for word pairs exhibiting different types of morphological relationships, including suffixed-stem (e.g., teacher-teach), suffixed-suffixed (e.g., saintly-sainthood), and prefixed-stem pairs (preheat-heat). The results can be understood in terms of connectionist models that use distributed representations rather than discrete morphemes. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Does English coronal place assimilation create lexical ambiguity?

The purpose of this study was to determine how potential lexical ambiguity produced by place assimilation is resolved. Four cross-modal form priming experiments using primes in sentential contexts were performed. In the first 2, prime items had underlyingly coronal offsets (e.g., right) with assimilated noncoronal place. The primes were judged to be perceptually ambiguous (between right and ripe) in Experiment 1 and noncoronal (ripe) in Experiment 2 in off-line testing. In Experiment 3, primes were replaced with corresponding underlyingly noncoronal items (ripe). In all 3 experiments, participants showed selective priming for the underlying form of the prime. A 4th form priming experiment using the gated tokens of priming stimuli used in Experiment 2 examined the role of postlexical context on this process. In this experiment, participants showed priming for both underlying and surface forms of the prime. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

跨模态交互融合与全局感知的RGB-D显著性目标检测

近年来, RGB-D显著性检测方法凭借深度图中丰富的几何结构和空间位置信息, 取得了比RGB显著性检测模型更好的性能, 受到学术界的高度关注. 然而, 现有的RGB-D检测模型仍面临着持续提升检测性能的需求. 最近兴起的Transformer擅长建模全局信息, 而卷积神经网络(CNN)擅长提取局部细节. 因此, 如何有效结合CNN和Transformer两者的优势, 挖掘全局和局部信息, 将有助于提升显著性目标检测的精度. 为此, 提出一种基于跨模态交互融合与全局感知的RGB-D显著性目标检测方法, 通过将Transformer网络嵌入U-Net中, 从而将全局注意力机制与局部卷积结合在一起, 能够更好地对特征进行提取. 首先借助U-Net编码-解码结构, 高效地提取多层次互补特征并逐级解码生成显著特征图. 然后, 使用Transformer模块学习高级特征间的全局依赖关系增强特征表示, 并针对输入采用渐进上采样融合策略以减少噪声信息的引入. 其次, 为了减轻低质量深度图带来的负面影响, 设计一个跨模态交互融合模块以实现跨模态特征融合. 最后, 5个基准数据集上的实验结果表明, 所提算法与其他最新的算法相比具有显著优势.     

基于模态语义增强的跨模态食谱检索方法

在跨模态食谱检索任务中,如何有效地对模态进行特征表示是一个热点问题。目前一般使用两个独立的神经网络分别获取图像和食谱的特征,通过跨模态对齐实现跨模态检索。但这些方法主要关注模态内的特征信息,忽略了模态间的特征交互,导致部分有效模态信息丢失。针对该问题,提出一种通过多模态编码器来增强模态语义的跨模态食谱检索方法。首先使用预训练模型提取图像和食谱的初始语义特征,并借助对抗损失缩小模态间差异;然后利用成对跨模态注意力使来自一个模态的特征反复强化另一个模态的特征,进一步提取有效信息;接着采用自注意力机制对模态的内部特征进行建模,以捕捉丰富的模态特定语义信息和潜在关联知识;最后,引入三元组损失最小化同类样本间的距离,实现跨模态检索学习。在Recipe 1M数据集上的实验结果表明,该方法在中位数排名(MedR)和前K召回率(R@K)等方面均优于目前的主流方法,为跨模态检索任务提供了有力的解决方案。     

成对相似度迁移哈希用于无监督跨模态检索

哈希编码能够节省存储空间、提高检索效率,已引起广泛关注.提出一种成对相似度迁移哈希方法(pairwise similarity transferring hash,PSTH)用于无监督跨模态检索.对于每个模态,PSTH将可靠的模态内成对相似度迁移到汉明空间,使哈希编码继承原始空间的成对相似度,从而学习各模态数据对应的哈希编码;此外,PSTH重建相似度值而不是相似度关系,使得训练过程可以分批进行;与此同时,为缩小不同模态间的语义鸿沟,PSTH最大化模态间成对相似度.在三个公开数据集上进行了大量对比实验,PSTH取得了SOTA的效果.     

耦合保持投影哈希跨模态检索

目的 基于哈希的跨模态检索方法因其检索速度快、消耗存储空间小等优势受到了广泛关注。但是由于这类算法大都将不同模态数据直接映射至共同的汉明空间,因此难以克服不同模态数据的特征表示及特征维度的较大差异性,也很难在汉明空间中同时保持原有数据的结构信息。针对上述问题,本文提出了耦合保持投影哈希跨模态检索算法。方法 为了解决跨模态数据间的异构性,先将不同模态的数据投影至各自子空间来减少模态“鸿沟”,并在子空间学习中引入图模型来保持数据间的结构一致性;为了构建不同模态之间的语义关联,再将子空间特征映射至汉明空间以得到一致的哈希码;最后引入类标约束来提升哈希码的判别性。结果 实验在3个数据集上与主流的方法进行了比较,在Wikipedia数据集中,相比于性能第2的算法,在任务图像检索文本（I to T）和任务文本检索图像（T to I）上的平均检索精度（mean average precision,mAP）值分别提升了6%和3%左右;在MIRFlickr数据集中,相比于性能第2的算法,优势分别为2%和5%左右;在Pascal Sentence数据集中,优势分别为10%和7%左右。结论 本文方法可适用于两个模态数据之间的相互检索任务,由于引入了耦合投影和图模型模块,有效提升了跨模态检索的精度。     

Failure to remap visuotactile space across the midline in the split-brain.

Examined the effect of posture change on the representation of visuotactile space in a 47-yr-old male split-brain patient using a cross-modal congruency task. The patient made speeded elevation discrimination responses (up vs down) to a series of tactile targets presented to the index finger or thumb of his right hand. The authors report congruency effects elicited by irrelevant visual distractors placed either close to, or far from, the stimulated hand. These cross-modal congruency effects followed the right hand as it moved within the right hemispace, but failed to do so when the hand crossed the midline into left hemispace. These results support claims that interhemispheric connections are required to maintain an accurate representation of visuotactile space. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Adaptation reveals multiple levels of representation in auditory stream segregation.

When presented with alternating low and high tones, listeners are more likely to perceive 2 separate streams of tones (“streaming”) than a single coherent stream when the frequency separation (Δ?) between tones is greater and the number of tone presentations is greater (“buildup”). However, the same large-Δ? sequence reduces streaming for subsequent patterns presented after a gap of up to several seconds. Buildup occurs at a level of neural representation with sharp frequency tuning. The authors used adaptation to demonstrate that the contextual effect of prior Δ? arose from a representation with broad frequency tuning, unlike buildup. Separate adaptation did not occur in a representation of Δ? independent of frequency range, suggesting that any frequency-shift detectors undergoing adaptation are also frequency specific. A separate effect of prior perception was observed, dissociating stimulus-related (i.e., Δ?) and perception-related (i.e., 1 stream vs. 2 streams) adaptation. Viewing a visual analogue to auditory streaming had no effect on subsequent perception of streaming, suggesting adaptation in auditory-specific brain circuits. These results, along with previous findings on buildup, suggest that processing in at least 3 levels of auditory neural representation underlies segregation and formation of auditory streams. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

改进U-Net3+与跨模态注意力块的医学图像融合

目的 针对目前多模态医学图像融合方法深层特征提取能力不足,部分模态特征被忽略的问题,提出了基于U-Net3+与跨模态注意力块的双鉴别器生成对抗网络医学图像融合算法(U-Net3+ and cross-modal attention block dual-discriminator generative adversal network,UC-DDGAN)。方法 结合U-Net3+可用很少的参数提取深层特征、跨模态注意力块可提取两模态特征的特点,构建UC-DDGAN网络框架。UC-DDGAN包含一个生成器和两个鉴别器,生成器包括特征提取和特征融合。特征提取部分将跨模态注意力块嵌入到U-Net3+下采样提取图像深层特征的路径上,提取跨模态特征与提取深层特征交替进行,得到各层复合特征图,将其进行通道叠加、降维后上采样,输出包含两模态全尺度深层特征的特征图。特征融合部分通过将特征图在通道上进行拼接得到融合图像。双鉴别器分别对不同分布的源图像进行针对性鉴别。损失函数引入梯度损失,将其与像素损失加权优化生成器。结果 将UC-DDGAN与5种经典的图像融合方法在美国哈佛医学院公开的脑部疾病图像数据集上进行实验对比,其融合图像在空间频率(spatial frequency,SF)、结构相似性(structural similarity,SSIM)、边缘信息传递因子(degree of edge information,Q^AB/F)、相关系数(correlation coefficient,CC)和差异相关性(the sum of the correlations of differences,SCD)等指标上均有提高,SF较DDcGAN(dual discriminator generation adversative network)提高了5.87%,SSIM较FusionGAN(fusion generative adversarial network)提高了8%,Q^AB/F较FusionGAN提高了12.66%,CC较DDcGAN提高了14.47%, SCD较DDcGAN提高了14.48%。结论 UC-DDGAN生成的融合图像具有丰富深层特征和两模态关键特征,其主观视觉效果和客观评价指标均优于对比方法,为临床诊断提供了帮助。