Previous research demonstrated that visual representations in working memory exhibit biases with respect to the categorical structure of the stimulus space. However, a majority of those studies used ...behavioral measures of working memory, and it is not clear whether the working memory representations per se are influenced by the categorical structure or whether the biases arise in decision or response processes during the report. Here, I applied a multivariate decoding technique to EEG data collected during working memory tasks to determine whether neural activity associated with the representations in working memory is categorically biased prior to the report. I found that the decoding of spatial working memory was biased away from the nearest cardinal location, consistent with the biases observed in the behavioral responses. In a follow-up experiment which was designed to prevent the use of a response preparation strategy, I found that the decoding still exhibited categorical biases. Together, these results provide neural evidence that working memory representations themselves are categorically biased, imposing important constraints on the models of working memory representations.
This study tested the hypothesis that even the simplest cognitive tasks require the storage of information in working memory (WM), distorting any information that was previously stored in WM. ...Experiment 1 tested this hypothesis by requiring observers to perform a simple letter discrimination task while they were holding a single orientation in WM. We predicted that performing the task on the interposed letter stimulus would cause the orientation memory to become less precise and more categorical compared to when the letter was absent or when it was present but could be ignored. This prediction was confirmed. Experiment 2 tested the modality specificity of this effect by replacing the visual letter discrimination task with an auditory pitch discrimination task. Unlike the interposed visual stimulus, the interposed auditory stimulus produced little or no disruption of WM, consistent with the use of modality‐specific representations. Thus, performing a simple visual discrimination task, but not a simple auditory discrimination task, distorts information about a single feature being maintained in visual WM. We suggest that the interposed task eliminates information stored within the focus of attention, leaving behind a WM representation outside the focus of attention that is relatively imprecise and categorical.
Pupillometry provides an accessible option to track working memory processes with high temporal resolution. Several studies showed that pupil size increases with the number of items held in working ...memory; however, no study has explored whether pupil size also reflects the quality of working memory representations. To address this question, we used a spatial working memory task to investigate the relationship of pupil size with spatial precision of responses and indicators of reliance on generalized spatial categories. We asked 30 participants (15 female, aged 19–31) to remember the position of targets presented at various locations along a hidden radial grid. After a delay, participants indicated the remembered location with a high‐precision joystick providing a parametric measure of trial‐to‐trial accuracy. We recorded participants' pupil dilations continuously during task performance. Results showed a significant relation between pupil dilation during preparation/early encoding and the precision of responses, possibly reflecting the attentional resources devoted to memory encoding. In contrast, pupil dilation at late maintenance and response predicted larger shifts of responses toward prototypical locations, possibly reflecting larger reliance on categorical representation. On an intraindividual level, smaller pupil dilations during encoding predicted larger dilations during late maintenance and response. On an interindividual level, participants relying more on categorical representation also produced larger precision errors. The results confirm the link between pupil size and the quality of spatial working memory representation. They suggest compensatory strategies of spatial working memory performance—loss of precise spatial representation likely increases reliance on generalized spatial categories.
A vast range of Natural Language Processing (NLP) systems that are in use today have direct impact on humans. While machine learning models are expected to automatically infer world knowledge from ...historical texts, we should also be cognizant not to let NLP applications consume undesired societal stereotype bias back. Many bias evaluation measures have been designed and experimented to check whether unwanted stereo-type bias is present in the model or not. Upon performing various experiments, we found out that the most popular bias measures do not always indicate bias accurately. In addition to these experimental findings, we also propose our novel Differential Cosine Bias measure with examples of unwanted stereotype biases as well as necessary categorical bias that is based on knowledge. Our experiments show that our bias measure is a potential indicator of bias in NLP models compared to the popular bias evaluation measures.
PDF
