아이들의 정신 회전 능력을 탐구하는 바이오 영감 기계학습 모델
📝 Abstract
Highlights -We use a bio-inspired machine learning models to examine evidence of mental rotation abilities in children -We show that a simple recognition strategy suffices to solve habituation-based tasks used to assess mental rotations in young children. -We demonstrate that a model forming expectations about a rotation transformation can explain children’s results in mental rotation tasks with the violation-of-expectation paradigm. -The investigated recognition strategies no longer work when the mental rotation tasks are made closer to adults'.
💡 Analysis
Highlights -We use a bio-inspired machine learning models to examine evidence of mental rotation abilities in children -We show that a simple recognition strategy suffices to solve habituation-based tasks used to assess mental rotations in young children. -We demonstrate that a model forming expectations about a rotation transformation can explain children’s results in mental rotation tasks with the violation-of-expectation paradigm. -The investigated recognition strategies no longer work when the mental rotation tasks are made closer to adults'.
📄 Content
Mental rotation (MR) refers to the mental transformation of one object view into another view of the object [57]. Performing MR is thought to be essential to recognize a complex shape from different viewpoints or to predict its appearance after a manual rotation. Thus, MR plays a key role in object perception and spatial reasoning. There is strong and diverse evidence for MR abilities in adults. In their seminal study [57], they presented adult participant with pairs of objects composed of 1-one view of a complex 3D object and 2-either the same object rotated in-depth or a mirror-reversed version. Importantly, only the 3D arrangement of basic features differs between the original and mirror versions of the object, limiting the cues available for recognition. The results showed that the recognition time was linearly dependent on the angle of rotation. After introspection, participants reported that they had to imagine an object as rotated in the same orientation as the original object to compare them. In general, these are strong arguments that support the interpretation that MR involves a kind of mental simulation of physical rotations.
Additional experiments have explicitly investigated the cognitive processes underlying MRs. These studies have shown that MR operates along the shortest angular path between two views [56], and that this holds across different tasks and stimuli [56]. Together with the well-established positive correlation between angular disparity and reaction time, this suggests that MR involves a holistic transformation through the shortest rotational trajectory separating two viewpoints. Further research has proposed that MR reflects a covert simulation of motor actions. In a seminal study, Wexler et al. [62] introduced a dual-task paradigm in which participants were required to perform an MR recognition task while simultaneously rotating a joystick at a fixed speed, either clockwise or counterclockwise. The results show that incongruence between the direction of manual rotation and the direction of MR impairs both recognition accuracy and response time, whereas congruence facilitates performance. Furthermore, several studies report similar response patterns between mental and physical object rotations [20,63], and observed that increasing the angular disparity between two views is associated with increased activations in the parietal and premotor brain regions [37,23]. Thus, in this paper, we assume that genuine MR involves a form of a stepwise mental transformation along a rotational path connecting two viewpoints.
Understanding the developmental origins of MR requires an assessment of the use of MR in young children. The current evidence for MR in infants and toddlers rests mainly on two behavioral paradigms based on looking time, adapted from adult protocols. These tasks are made much easier by simplifying objects and providing additional cues to help children. For instance, these include a broader range of orientations of an object [39] or indirect information about the rotation applied between habituation and test images [19]. In addition, evidence of object recognition mostly relies on differences in looking time between an original object and its mirror version, without evidence for a linear relation between looking time and the angle of rotation. This is crucial as object recognition alone may also be the outcome of mental processes different from genuine MR [60,10,25,7]. For instance, children may simply compare object views based on low-level features like object apparent size or features that remain constant over different orientations. To test MR abilities of children from 1 to 3 years old with similarly complex objects, a recent staircase procedure asks children to solve a rotation task with eye movements [4]. However, there is no evidence of a linear relation between reaction time and rotation angle. Overall, it remains unclear whether children rely on genuine MR to solve these tasks.
In this work, we use computational models to explore the role of simple object recognition strategies in solving standard MR tasks for children. We simulate three classical MR tasks based on 1-children’s habituation; 2-adults’ recognition and 3-children’s violation-of-expectation. For these tasks, we model the recognition process with a recent bio-inspired machine learning model that learns visual representations [2]. It is trained by simulating child locomotion training, which is known to improve infants’ abilities to solve these tasks [16,54,53,55,38,19]. After training, the vision model can simulate two recognition strategies used by humans. The first strategy [5,36,59] simulates view-invariant exemplar-based object recognition by comparing the similarity of representations of different views without further mental processing. The second recognition process predicts the rotation transformation between two viewpoints of an object. Importantly, we prevent the model from using implicit form
This content is AI-processed based on ArXiv data.