Past Research Projects
Object Learning with Simultaneous Presentation in Humans and Deep Neural Networks
This follow-up study for our previous experiment examined whether the observed human full-set disadvantage is due to working memory limitations. We adopted simultaneous presentation to mitigate the memory demand and keep other experiment designs the same as previous experiments.
Comparing Human Continual Learning with Deep Neural Networks
This project compares human continual learning with deep neural networks in classification (or categorization) tasks. One of the main challenges in machine learning is that the network readily forgets previously learned information as it encounters new information. For example, a deep network pre-trained with some images will acquire useful features for tasks such as classification. However, when training the network with more images from other sources to further improve its performance or to perform new tasks, it will show complete forgetting of the previously learned tasks, which is called catastrophic forgetting. In collaboration with Prof. Derek Hoiem and Zhen Zhu at the Department of Computer Science, we examined human and machine category learning under the same scenario to see whether and why the learning sequence affects humans and neural networks differently.
Examining the Linear Separability Effect in Color Space
(Bachelor’s Thesis)
Using color stimuli defined in CIELab color space, D'Zmura (1991) found a linear separability effect in the heterogeneous visual search task. When two types of distractors could be linearly separated from the target in this color space, the search task was easy; otherwise, the search would be difficult (Treisman & Gelade, 1980; Bauer et al., 1996). Although the linear separability effect is verified in multiple feature spaces and is believed to determine search efficiency, there are still some doubts about its existence (Lindsey et al., 2010; Vighneshvel and Arun, 2013; Wolfe, 2021). To examine the existence of the linear separability effect in color space, we adopted the same predictive approach utilized by Xu et al. (2021) when examining the linear separability effect in orientation space.
How Navigation Information Affects Perceptions of Time and Space in Virtual Urban Environments
To study how the navigation information on signs influences human perception of time and space in the street environment, we placed human subjects at one end of a virtual urban street corridor with different amounts of sign information at each intersection. Due to the pandemic, we transferred this project online by PsychoPy and Pavlovia. The subjects would be presented with a 3D street environment created by Unity and observe the movement to the destination. Questions were given randomly to ask subjects about past time and traveled distance.