基于名人面孔视觉特征和语义信息的视觉统计学习

Visual Statistical Learning Based on the Visual Feature and Semantic Information of Famous Faces

视觉统计学习是指个体依据视觉刺激之间的转接概率来掌握统计规律的过程。本研究通过5个实验探讨了个体基于名人面孔视觉特征和语义信息进行视觉统计学习的加工机制。每个实验均包括熟悉(学习)和测试两个阶段:在熟悉阶段,让被试观看名人面孔并完成重复图片探测的无关任务;在测试阶段,让被试进行二选一迫选任务。其中,实验1和2分别考察基于名人面孔视觉特征和语义信息的视觉统计学习效果;实验3分别考察基于名人面孔视觉特征和语义信息视觉进行统计学习的精确性;实验4进一步考察基于名人面孔视觉特征和语义信息进行视觉统计学习的时间特征;实验5验证基于名人面孔视觉特征的视觉统计学习具有面孔特异性。结果表明:个体能同时基于名人面孔视觉特征和语义信息进行精确的视觉统计学习;基于正立名人面孔的视觉统计学习效果显著高于基于倒置名人面孔的视觉统计学习效果;虽然基于视觉特征和语义信息的统计加工都具有一致的精确性,但后者需要更多的加工时间。这提示:基于名人面孔视觉特征的视觉统计学习具有面孔特异性,个体基于名人面孔视觉特征和语义信息的视觉统计学习过程是分离的,统计运算发生于面孔特征加工完成之后。

The environment contains considerable information distributed across time and space, and the operation of visual statistical learning（VSL） allows our visual system to be remarkably sensitive to them. VSL can be defined as the ability of human observers to extract information about the joint and conditional probabilities of stimuli co-occurring. Previous studies have already found two distinctive types of VSL： one that based on visual feature, and the other that based on semantic information. However, people learn to obtain novel statistical regularities based on human faces in the real world, yet mere use of novel shapes or objects was found in those experiments. To our knowledge, none of the existing research has addressed VSL based on human faces. In the present study, combined use of VSL paradigm and backward test technique was thereby employed to investigate the extent to which participants＇ computation of statistical regularities was based on the visual features and semantic information of famous faces.We conducted four experiments, and each included a familiarization stage and a familiarity test stage. During the familiarization stage, the participants watched a 20-min sequence of 1,248 face pictures（include 96 cover task pictures）, presented one at a time for 300 ms and with a 700ms（or 400ms） inter-stimulus interval（ISI）. Each face was assigned a unique position within one of four triplets（e.g., A-B-C-G-H-I-D-E-F-J-K-L…）. The（cover） task was to detect back-to-back repeats of the same picture（96 trials） and to indicate them as quickly as possible by hitting the space bar. Following this stage, the participants received a surprise two-interval forced-choice（2IFC） familiarity test. On each test trial, they viewed two 3-image test sequences, a familiar triplet（e.g., A-B-C） and a foil triplet（e.g., A-E-I）, which were segmented from each other by an additional 1,000-ms pause. The participants were asked to indicate whether the first or second test sequence seemed more familiar based on the familiarization stage. We examined whether VSL can operate on statistical regularities of famous faces with face pictures in the test stage of Experiment 1 and with name stimuli representing each face in the test stage of Experiment 2. In experiment 3, the test stage contained both the forward triplets（e.g., A-B-C） and backward triplets（e.g., C-B-A） to examine whether temporal order information is critical to the operation of VSL based on famous faces. In Experiment 4, the ISI of face pictures in familiarization stage was reduced（from 700 ms to 400ms） to investigate the time attributes of VSL based on visual features and semantic information of famous faces. In experiment 5, we manipulated face display orientation（upright vs. inverted） to investigate the specificity of face in VSL. The results of Experiment 1 and 2 showed that in the familiarity task（the measure of VSL）, performance was markedly above chance levels with visual feature and name stimuli of famous faces. In Experiment 3, participants showed significant VSL only in the forward triplets test condition with both the visual feature and name stimuli of famous faces. It was noteworthy that a significant learning effect was found in experiment 4A, where the stimuli were famous face pictures. Contrary to results of experiment 4A, VSL was not observed in the name test condition（experiment 4B）. In experiment 5, the performance for upright faces was significantly better than inverted faces in VSL. In summary, the present results provide clear evidence that the VSL based on visual features and semantic information of famous face is face specific. Based on the time procedure, it suggests processes of visual features and semantic information of famous face in VSL are separate, and the statistical computation of the temporal order is successively behind the face processing.