Bogeng Song

Coming soon...

In the current study, we designed a motion coherence threshold task to measure the visual discrimination of coherent motion stimuli that are accelerated or decelerated by natural gravity (1g) or reversed gravity (-1g). Using the QUEST threshold method, we measured the proportion of signal dots required for participants to discriminate the direction of upward or downward coherent motion, i.e., perceptual threshold, under different gravity conditions.

(Not final version) To preserve one's self-respect, our memory of unethical actions gradually become obfuscated, a cognitive process known as unethical amnesia. This self-serving dishonesty may change human memory from two aspects: memory accuracy and metacognitive confidence. To test this, we conducted two fMRI studies (the second a pre-registered replication of the first) to identify how repeated dishonest responses over time distorts memory accuracy and confidence. These included fMRI in an information passing task, and pre-task memory and post-task memory testing combined with mouse-tracking. Our study shows three main findings: (1) memory decreased for xx condition, RESULTING IN CHANGES IN ACTIVITY IN THE XXX regions of xxx ; (2) people show adaptation in the moral decisions, but with xxx, indicated by the DDM model; (3) the ISRSA???? results provided further evidence that more xxx , indicating the individual difference in both the moral decision and its after-effect on memory. We find that the accuracy decreases when the XXX even occurs for individuals whose choice is self-interests driving. Crucially, these motivated memory decreases are irrespective of the memory confidence and this effect disappears under conditions of random dishonesty. Together, these findings suggest that memory changes in unethical decisions are XXXX.

To test unethical amnesia behavior, we conducted two fMRI studies to identify how repeated dishonest responses over time distort memory accuracy and confidence. In this project, I mainly analyze behavioral data and try to model fMRI and mouse tracking data to predict the time-series subjects' memory accuracy change. Now, I'm basically done with the data analysis part, and I'm summarizing the results, doing paper review, and writing the paper.

Coming soon...

In this project, I participated in the whole process except the data collection part.

Coming soon...

In this project, I ran several deep learning model to deal with fMRI images, create figures and write paper.

Previous studies investigating working memory performance in patients with mild cognitive impairment (MCI) have mainly focused on the neural mechanisms of alterations in activation. To date, very few studies have investigated background network alterations in the working memory state. Therefore, the present study investigated the static and temporal dynamic changes in the background network in MCI patients during a working memory task. A hybrid delayed-match-to-sample task was used to examine working memory performance in MCI patients. Functional magnetic resonance imaging (fMRI) data were collected and the marker of amplitude of low-frequency fluctuations (ALFF) was used to investigate alterations in the background network. The present study demonstrated static and dynamic alterations of ALFF in MCI patients during working memory tasks, relative to the resting state. Traditional static analysis revealed that ALFF decreased in the right ventrolateral prefrontal cortex (VLPFC), right dorsolateral PFC (DLPFC), and left supplementary motor area for normal controls (NCs) in the working memory state. However, the same regions showed increased ALFF in MCI patients. Furthermore, relative to NCs, MCI patients demonstrated altered performance-related functional connectivity (FC) patterns, with the right VLPFC and right DLPFC as ROIs. In terms of temporal dynamic analysis, the present study found that in the working memory state dynamic ALFF of bilateral thalamus regions was increased in NCs but decreased in MCI patients. Additionally, MCI patients demonstrated altered performance-related coefficient of variation patterns; the regions in MCI patients were larger and more widely distributed in the parietal and temporal lobes, relative to NCs. This is the first study to examine static and temporal dynamic alterations of ALFF in the background network of MCI patients in working memory states. The results extend previous studies by providing a new perspective on the neural mechanisms of working memory deficits in MCI patients.

In this project, I analyze fMRI data and write the paper results part.

Microsaccades, or small recurring eye movements, typically occur ~1-2 times per second. Although generally considered involuntary, the characteristics of these eye movements are task-dependent and affect performance. For example, microsaccades are flexibly allocated to precisely relocate gaze during high acuity tasks in the fovea and suppressed prior to the onset of a stimulus outside of the fovea during a motion discrimination task. Here we investigated whether and how microsaccade rates are adaptively modulated by trial difficulty when observers discriminate motion directions in the perifovea. METHODS. We used a 2AFC task to measure the discriminability of a Gabor drifting for 500ms in 1 of 8 reference directions (4 cardinal, 4 oblique) at 8 isoeccentric locations (7​​°). Observers reported a Gabor’s drift direction, which was slightly clockwise or counterclockwise with respect to a reference direction. The difficulty of each trial was varied in two ways: (a) cardinal vs. oblique directions; and (b) tilt offset between reference and Gabor direction. The tilt angles were randomized using a method of constant stimuli: the Gabor’s drift direction was offset from the reference direction by ± 0.5, 1, 2, 4 or 8°. RESULTS. First, we found that microsaccade rates were suppressed prior to stimulus onset. Second, microsaccade rates were modulated by trial difficulty in two ways: they decreased (1) as the angular offset between the target and standard decreased; and (2) when stimuli drifted in oblique rather than cardinal directions. CONCLUSION. Microsaccades were suppressed prior to stimulus presentation, and during the stimulus period for difficult discrimination tasks in the perifovea. This flexibility is consistent with the proposals that greater fixational stability can (1) mitigate potential blur during microsaccades, and (2) prolong the duration of evidence accumulation. As a result, microsaccades may serve as a marker of cognitive effort for visual tasks in the perifovea.

Delineating visual field maps and iso-eccentricities from fMRI data is an important but time-consuming task for many neuroimaging studies on the human visual cortex because the traditional methods of doing so using retinotopic mapping experiments require substantial expertise as well as scanner, computer, and human time. Automated methods based on gray matter anatomy or a combination of anatomy and functional mapping can reduce these requirements but are less accurate than experts. Convolutional Neural Networks (CNNs) are powerful tools for automated medical image segmentation. We hypothesize that CNNs can define visual area boundaries with high accuracy. We trained U-Net CNNs with ResNet18 backbones to predict either V1, V2, and V3 boundaries or 5 regions of iso-eccentricity using human-labeled maps. Separate CNNs were trained to predict these regions using different combinations of the following input data: (1) anatomical data from a T1-weighted image only, (2) anatomical data from T1-weighted and T2*-weighted images, (3) white-matter tract endpoints from diffusion-weighted imaging, (4) functional data from retinotopic mapping. All CNNs using functional data had cross-validated accuracy that was statistically indistinguishable from the inter-rater reliability of the training dataset (dice coefficient of 92%) while the CNNs lacking functional data had lower but similar accuracies (~75%). Existing models that do not use CNNs had accuracies lower than any of the CNNs. These results demonstrate that with current methods and data quality, CNNs can replace the time and effort of human experts in manually defining early retinotopic maps, but cannot yet replace the acquisition of functional data.

Use afni and SPM to analyze fMRI data, use psychopy to write experimental programs, basic data analysis and organization.

We used psychophysical and eye-tracking methods to measure behavioral differences for different motion directions (e.g., radial versus tangential). To model the behavioral data, we fit psychophysical curves, and compare performance for different motion directions.

Our memory is an integral centrepiece in the process of storytelling however, our intuition leads us to believe that we use different types of memory and memory retrieval for different types of stories. To further examine the relationships, we built a framework of different computational models to better understand the cognitive processes that people use while constructing a type of story. Our primary dataset, hippoCorpusV2, contains 6,854 diary-like short stories individually labeled into three categories: recall, imagine and retold. We used relevantMachine Learning models and techniques to classify these three types of stories, and extract the corresponding features to compare with our conclusions from human behavior experiments to better understand people's cognitive processes. We found that our model results and behavioral results are similar, and there are three main characteristics that help us distinguish the three types of stories: the time it takes to build the story, the amount of concrete, specific events mentioned in a story, and detailed, sensory information providing background color to the story. From these results, we can infer that recall stories are based on the direct retrieval of episodic memory, while imagined stories are mainly generated based on general knowledge of semantic memory. While retold stories do retain some details in episodic memory, they also require general knowledge due to the inherent human tendency to forget trivial details.

Existing work on Semi-Supervised Learning with Variational Bayesian Inference and Maximum Uncertainty Regularization has shown clear improvements in classification errors of various Consistency Regularization based methods. Functional Space Variational Inference is an improvement to Variational Inference. We propose a method combining Functional Space Variational Inference and Consistency Regularization by minimizing the KL divergence of distributions over functions. We apply our method to the partially labeled datasets and compare the three ways to realize our method.

Increasing research has shown that objects associated with rewards can effectively capture attention. Perceptual load theory believes that under low load conditions, the current task just uses only a proportion of attention resource, the spared resource will automatically spread to distractor ; but under high perceptual load conditions, all attention resources are consumed, and no extra attention resources are available to process interference stimuli. However, it is not known that after the interference stimulus and the reward are established, the interference stimulus associated with the reward can effectively capture the attention resource under high and low load conditions. This study aims to explore this issue in conjunction with the learning-test paradigm and ERP techniques. During the experiment, the subjects were asked to link the colors (red, green) to the high and low rewards during the learning task. In the test task, the handy experimental paradigm was used to explore the centrality by setting the interference stimuli and perceived load levels of different colors. The amplitude of the P1 band around the concave changes. We assume that regardless of the level of load, the interfering stimulus associated with the high reward will always produce a corrected P1 amplitude. Behavioral results show significant differences in the correct rates for high load and low load conditions. The results of EEG showed that in the occipital cortex, the reward was significant. Compared with the low reward condition, the P1 amplitude was higher under the high reward condition; and the interaction between the reward and the perceptual load was significant, and the simple effect test showed that it was high or low. Under the perceptual load condition, there is a significant difference in the P1 amplitude caused by the stimulation of the high and low rewards, and in the case of low perceptual load, the difference between the high and low rewards is significantly greater than the difference between the high and low rewards under the high perceptual load. According to the results, even under high-perceptual load conditions, the interfering stimulus associated with the high reward still captures attention. Attention to irrelevant stimuli captures a common adjustment of the value-driven attentional effect and the perceived load level of the current task.