Publications

Despite their impressive performance in object recognition and other tasks under standard testing conditions, deep networks often fail to generalize to out-of-distribution (o.o.d.) samples. One cause for this shortcoming is that modern architectures tend to rely on "shortcuts" - superficial features that correlate with categories without capturing deeper invariants that hold across contexts. Real-world concepts often possess a complex structure that can vary superficially across contexts, which can make the most intuitive and promising solutions in one context not generalize to others. One potential way to improve o.o.d. generalization is to assume simple solutions are unlikely to be valid across contexts and avoid them, which we refer to as the too-good-to-be-true prior. A low-capacity network (LCN) with a shallow architecture should only be able to learn surface relationships, including shortcuts. We find that LCNs can serve as shortcut detectors. Furthermore, an LCN's predictions can be used in a two-stage approach to encourage a high-capacity network (HCN) to rely on deeper invariant features that should generalize broadly. In particular, items that the LCN can master are downweighted when training the HCN. Using a modified version of the CIFAR-10 dataset in which we introduced shortcuts, we found that the two-stage LCN-HCN approach reduced reliance on shortcuts and facilitated o.o.d. generalization.

Incidental findings defined as valuable findings that are not searched purposely by the experts were originally discovered by radiologists. Despite the importance and great practical value of this phenomenon for visual search, it was almost not studied by cognitive psychologists and vision science experts. The current study aimed to examine experimentally incidental findings in visual search. The main objective was to clarify independence of incidental findings from subsequent search misses, another well-known visual search phenomenon. In order to do that, the standard experimental paradigm for detecting subsequent search misses was used. At the same time the stimuli material and tasks were created to closely fit the definition of incidental findings. The findings revealed that subsequent search misses, but not incidental findings were observed. The results suggest that incidental findings may be closely related to subsequent search misses. As well as that, the difficulty of the task, particularly, induced by target-distractor similarity, may be one of the major factors, leading to the emergence of subsequent search misses instead of incidental findings.

Subsequent search misses (SSM) refer to the decrease in accuracy of second target detection in dual-target visual search. One of the theoretical explanations of SSM errors is similarity bias – the tendency to search for similar targets and to miss the dissimilar ones. The current study focuses on both perceptual and categorical similarity and their individual roles in SSM. Five experiments investigated the role of perceptual and categorical similarity in subsequent search misses, wherein perceptual and categorical similarities were manipulated separately, and task relevance was controlled. The role of both perceptual and categorical similarity was revealed, however, the categorical similarity had greater impact on second target detection. The findings of this research suggest the revision of the traditional perceptual set hypothesis that mainly focuses on perceptual target similarity in multiple target visual search.

The disfluency effect is an improvement in the retention of information if its processing is difficult. Disfluency is usually attained by changing perceptual characteristics, and in the case of texts, fonts are made less legible. The disfluency effect has been demonstrated in several studies, but the number of unsuccessful replications is significantly higher. Among the possible reasons why the disfluency effect often cannot be detected, we consider the influence of other factors acting as moderators. This study examines the presence of images accompanying text information as a moderator. Based on dual-coding theory, we assume that the disfluency effect will manifest itself only when there are no images, since their presence contributes to better memorization. Two experiments were conducted in which information written either in Arial (fluent condition) or Comic Sans (disfluent condition) was presented to participants. The information was also accompanied by images or not. In the first experiment, the participants were asked to memorize 40 short words, in the second 7 fictional facts about the Earth. In both cases, the hypotheses were not confirmed—the disfluency effect did not manifest under any conditions.

The disfluency effect was recently discovered, according to which it is possible to improve memorization by making fonts less legible. In a number of studies, the existence of this effect has been discussed, and the question about the possible existence of moderators for this effect has also been raised. In this paper, the presence of images accompanying text information was considered as a moderator for the disfluency effect. 40 words were shown to study participants; depending upon the experimental group, the presented words were written in either a fluent or disfluent font, and accompanied by images or not. A significant improvement in memory was found in the presence of images, but the disfluency effect was not detected, and the presence of images did not moderate this effect.