A red code triggers an unintended approach motivation toward sweet ultra-processed foods: Possible implications for front-of-pack labels

Front-of-package labels (FOPL) are recommended to reduce consumer intake of ultra-processed food products (UPP). The multiple traffic-light label is one example of FOPL that indicates the content of target nutrients in products by displaying red (high), amber (medium), and/or green (low) color-coding. The red code may implicitly enhance sweetness perception and approach dispositions toward sweet UPP via cross-modal visual-taste interactions. We conducted two experiments to examine the possibility of contradictory influence of explicitly learned and implicit cross-modal associations on the emotional responses evoked by UPP pictures. In both experiments, we first explicitly associated the color codes with health-related meanings. In Experiment I (n = 78), a psychometric tool estimated the emotional responses (pleasantness and arousal ratings) evoked by UPP pictures when preceded by red, amber, or green color-codes. In Experiment II (n = 24), we recorded participants’ electrocortical brain activity to assess the early posterior negativity (EPN) component as an index of the emotional responses to UPP. The reported pleasantness (Experiment I) and the EPN amplitude (Experiment II) were greater for sweet UPP relative to salty UPP when primed with red codes but not when primed with green or amber. A red code increased positive emotions toward sweet UPP despite its explicit association with increased health-risks. Thus, the use of multiple traffic-lights might lead to an unintended implicit approach behavior toward sweet UPP. Designers, researchers, and policy makers may consider color-taste cross-modal associations when designing, testing, and applying FOPL.     

Reciprocal bi-directional generative adversarial network for cross-modal pedestrian re-identification

To improve the accuracy of cross-modal pedestrian re-identification,a reciprocal bi-directional generative adversarial network-based method is proposed.First,we build two generative adversarial networks to generate cross-modal heterogeneous images.Second,an associated loss is designed to pull close the distribution of features in latent space during the image translation between visible and infrared images so as to help the networks generate fake heterogeneous images that have high similarity with the real images.Finally,by concatenating the original and generated heterogeneous pedestrian images into the discriminative feature extraction network,images from different modalities can be unified into a common modality,thus suppressing the cross-modal gap.Representation learning and metric learning are utilized to achieve more discriminative pedestrian features.Comparative experiments are conducted on SYSU-MM01 and RegDB datasets to analyze the accuracy with different loss functions.Compared with other state-of-the-art cross-modal pedestrian re-identification methods,the proposed method achieves a higher accuracy and stronger robustness.     

基于深度学习的跨模态检索综述

由于多模态数据的快速增长,跨模态检索受到了研究者的广泛关注,其将一种模态的数据作为查询条件检索其他模态的数据,如用户可以用文本检索图像或/和视频。由于查询及其检索结果模态表征的差异,如何度量不同模态之间的相似性是跨模态检索的主要挑战。随着深度学习技术的推广及其在计算机视觉、自然语言处理等领域的显著成果,研究者提出了一系列以深度学习为基础的跨模态检索方法,极大缓解了不同模态间相似性度量的挑战,本文称之为深度跨模态检索。本文从以下角度综述有代表性的深度跨模态检索论文,基于所提供的跨模态信息将这些方法分为3类：基于跨模态数据间一一对应的、基于跨模态数据间相似度的以及基于跨模态数据语义标注的深度跨模态检索。一般来说,上述3类方法提供的跨模态信息呈现递增趋势,且提供学习的信息越多,跨模态检索性能越优。在上述不同类别下,涵盖了7类主流技术,即典型相关分析、一一对应关系保持、度量学习、似然分析、学习排序、语义预测以及对抗学习。不同类别下包含部分关键技术,本文将具体阐述其中有代表性的方法。同时对比提供不同跨模态数据信息下不同技术的区别,以阐述在提供了不同层次的跨模态数据信息下相关技术的关注点与使用异同。为评估不同的跨模态检索方法,总结了部分代表性的跨模态检索数据库。最后讨论了当前深度跨模态检索待解决的问题以及未来的研究方向。     

基于特征融合的无监督跨模态哈希

已有的无监督跨模态哈希（UCMH）方法主要关注构造相似矩阵和约束公共表征空间的结构,忽略了2个重要问题：一是它们为不同模态的数据提取独立的表征用以检索,没有考虑不同模态之间的信息互补;二是预提取特征的结构信息不完全适用于跨模态检索任务,可能会造成一些错误信息的迁移。针对第一个问题,提出一种多模态表征融合结构,通过对不同模态的嵌入特征进行融合,从而有效地综合来自不同模态的信息,提高哈希码的表达能力,同时引入跨模态生成机制,解决检索数据模态缺失的问题;针对第二个问题,提出一种相似矩阵动态调整策略,在训练过程中用学到的模态嵌入自适应地逐步优化相似矩阵,减轻预提取特征对原始数据集的偏见,使其更适应跨模态检索,并有效避免过拟合问题。基于常用数据集Flickr25k和NUS-WIDE进行实验,结果表明,通过该方法构建的模型在Flickr25k数据集上3种哈希位长检索的平均精度均值较DGCPN模型分别提高1.43%、1.82%和1.52%,在NUS-WIDE数据集上分别提高3.72%、3.77%和1.99%,验证了所提方法的有效性。     

跨模态表征与生成技术

多媒体数据持续呈现爆发式增长并显现出异源异构的特性,因此跨模态学习领域研究逐渐引起学术和工业界的关注。跨模态表征与生成是跨模态学习的两大核心基础问题。跨模态表征旨在利用多种模态之间的互补性剔除模态之间的冗余,从而获得更为有效的特征表示;跨模态生成则是基于模态之间的语义一致性,实现不同模态数据形式上的相互转换,有助于提高不同模态间的迁移能力。本文系统地分析了国际与国内近年来跨模态表征与生成领域的重要研究进展,包括传统跨模态表征学习、多模态大模型表示学习、图像到文本的跨模态转换和跨模态图像生成。其中,传统跨模态表征学习探讨了跨模态统一表征和跨模态协同表征,多模态大模型表示学习探讨了基于Transformer的模型研究,图像到文本的跨模态转换探讨了图像视频的语义描述、视频字幕语义分析和视觉问答等领域的发展,跨模态图像生成从不同模态信息的跨模态联合表示方法、图像的跨模态生成技术和基于预训练的特定域图像生成阐述了跨模态生成方面的进展。本文详细综述了上述各个子领域研究的挑战性,对比了国内外研究方面的进展情况,梳理了发展脉络和学术研究的前沿动态。最后,根据上述分析展望了跨模态表征与生成的发展趋势和突破口。     

以图像为主的多模态感知与多源融合技术发展及应用综述

不同模态的表现方式不同,描述目标的角度也会不同。多模态感知与多源融合技术是将两种或两种以上的模态组合起来,融合不同传感器、不同平台收集到的数据、特征信息,兼顾不同图像的优势,在遥感监测、军事安防、自动驾驶等领域有着广泛的运用。介绍了热成像、高光谱成像、偏振成像、合成孔径雷达(SAR)、激光雷达(LiDAR)等多模态感知技术,总结了不同成像方式的特点与联系,简述了多源融合相关概念及其技术发展历程,重点分析了不同模态下图像融合案例,在此基础上归纳多模态感知和多源融合技术的发展趋势,最后基于融合算法、系统整体性、评价指标总结出进一步发展方向。     

融合多重实例关系的无监督跨模态哈希检索

大多数跨模态哈希检索方法仅使用余弦相似度进行特征匹配,计算方式过于单一,没有考虑到实例的关系对于性能的影响.为此,提出一种基于多重实例关系图推理的方法,通过构造相似度矩阵,建立全局和局部的实例关系图,充分挖掘实例之间的细粒度关系.在多重实例关系图的基础上进行相似度推理,首先分别进行图像模态和文本模态关系图内部的推理,然后将模态内的关系映射到实例图中进行推理,最后执行实例图内部的推理.此外,为了适应图像和文本两种模态的特点,使用分步训练策略训练神经网络.在MIRFlickr和NUS-WIDE数据集上实验表明,提出的方法在mAP指标上具有很明显的优势,在Top-k-Precision曲线上也获得良好的效果.这也说明所提方法对实例关系进行深入挖掘,从而显著地提升检索性能.     

The Challenges of Cross-Modal Translation: English-to-Sign-Language Translation in the Zardoz System

The sign languages used by deaf communities around the world represent a linguistic challenge that natural-language researchers in AI have only recently begun to take up. This challenge is particularly relevant to research in Machine Translation (MT), as natural sign languages have evolved in deaf communities into efficient modes of gestural communication, which differ from English not only in modality but in grammatical structure, exploiting a higher dimensionality of spatial expression. In this paper we describe Zardoz, an on-going AI research system that tackles the cross-modal MT problem, translating English text into fluid sign language. The paper presents an architectural overview of Zardoz, describing its central blackboard organization, the nature of its interlingual representation, and the major components which interact through this blackboard both to analyze the verbal input and generate the corresponding gestural output in one of a number of sign variants.     

Stress Versus Coarticulation: Toward an Integrated Approach to Explicit Speech Segmentation.

Although word stress has been hailed as a powerful speech-segmentation cue, the results of 5 cross-modal fragment priming experiments revealed limitations to stress-based segmentation. Specifically, the stress pattern of auditory primes failed to have any effect on the lexical decision latencies to related visual targets. A determining factor was whether the onset of the prime was coarticulated with the preceding speech fragment. Uncoarticulated (i.e., concatenated) primes facilitated priming. Coarticulated ones did not. However, when the primes were presented in a background of noise, the pattern of results reversed, and a strong stress effect emerged: Stress-initial primes caused more pruning than non-initial-stress primes, regardless of the coarticulatory cues. The results underscore the role of coarticulation in the segmentation of clear speech and that of stress in impoverished listening conditions. More generally, they call for an integrated and signal-contingent approach to speech segmentation. (PsycINFO Database Record (c) 2010 APA, all rights reserved)