Research

Imagine you're driving and sitting in the red car. There’s a black car ahead of you in the same lane. Even if the actual distance between the two cars is identical for everyone, different people might perceive it differently. One person might see the black car as very close, while another might see it as farther away.

Even if their judgments about distance are similar, their confidence in those judgments might differ—and that affects how they behave. Someone who is confident that the car is close will likely reduce speed. But if they’re uncertain, they might simply maintain their speed. Now, imagine someone perceives the car as far away. If they’re confident in that judgment, they’ll likely continue driving without slowing down. But if they’re unsure, they might still decide to reduce—just in case.

So what does this tell us? The same decision—to keep speed or not—can arise from very different internal processes. What we’re seeing here is a combination of both sensory and decisional biases working together to shape behavior.

1) The perceptual biases induced by adaptation were not only stronger but also had a much slower buildup and decay in the peripheral compared to the central visual field, indicating that peripheral vision might be qualitatively different from central vision (Gao et al., 2019, Journal of Vision). 

2) Both the contrast threshold and adaptive biases were much larger for the elderly compared to the young. However, when the adaptive biases were normalized by the threshold, I found broader orientation tuning (i.e., larger threshold elevation after adapting to stimuli of orthogonal orientations) but not changes of dynamics with aging (Gao et al., 2021, Vision Research).

3) A period of 8 seconds per face was required to generate significant face gender aftereffects, suggesting a potentially slower mechanism of adaptation for face gender (Gao et al., 2022, Journal of Vision) compared to colors. When the total adaptation duration was the same, the aftereffects to repeat presentation of a constant face (alternated with a uniform gray screen) was the same regardless of the temporal frequency of the face presentation, indicating that adaptative biases for faces depends on total exposure duration but not exposure frequency. 

4) Our visual system learned the “adapted state”: aftereffects for contrast and motion were reduced after repeated adaption (or “training”) over days, and this “learned” state persists as long as at least a few months (Dong, Gao et al, 2016, Scientific Reports, co-first author). The findings indicate that our brain actively adjusts and reduce the perceptual biases to overcome the perceptual catastrophe.

FFDE is a great example that nicely illustrates our perception doesn’t really match the physical world. Let’s try it together. Please keep your eyes on the center cross. The distortion exists only in your perceptual experience, not in the images themselves. The phenomenon that continuously presents normal faces, especially in the peripheral visual field, which causes strong distortion, is called the Flashed Face Distortion illusion. Traditionally, researchers use subjective rating methods to measure perceptual illusions—especially complex ones like the Flashed Face Distortion Effect. This typically involves asking participants to rate how distorted the faces appear to them. However, to fully understand such complex illusions, it’s not enough to simply measure how strong the illusion feels based on såbjective ratings. The ultimate goal is to objectively quantify the illusion so that we can uncover the underlying neural mechanisms. This level of insight cannot be achieved through traditional rating methods alone. 

To dissociate the effects of decisional biases from sensory biases in this strong illusion, I developed a novel paradigm that objectively quantifies the illusion. I included two conditions in the experiment. In the illusion condition, I presented 15 flashes of faces to both sides of the fixation; these trials are expected to induce a strong illusion. In the control condition, I only presented a single flash of faces. Critically, to objectively quantify the illusion, on half of the trials in both conditions, I will replace one of the last two faces with a physically distorted face. Here are sample distorted faces using different distortion methods for the same face identity. At the end of each trial, participants were asked is either of the last two faces was physically distorted. The idea is that if people are subjectively experiencing the face distortion illusion, then the ‘physically non-distorted’ faces should be perceptually indistinguishable from those that are physically distorted.

Using this objective method, I found that decisional biases gradually increased over time to counteract the strong illusion (Gao, Chen & Rahnev, 2024, Psychonomic Bulletin and Review).

In addition, we also measured how FFDE is determined by face inversion, presentation rate, and the temporal gap between successive faces (Gao, Wang & Rahnev, 2024, Cognition).

In most cases, confidence and accuracy go hand in hand, however, in order to examine to what extent they can actually determine what we perceive, we need to dissociate the two.

My work investigated whether automatic multisensory integration follows objective performance or subjective confidence reports. We created two types of visual stimuli: high- vs. low-energy visual stimuli. The high-energy stimuli had higher confidence but worse performance, whereas the low-energy stimuli had lower confidence but better performance. We found that the low-energy visual stimuli had a larger impact and was given a larger weight on auditory motion judgement. By assuming a model where visual and auditory signals are combined differently without normalizations, my work reveals common computations between automatic multisensory integration and metacognitive confidence reports, suggesting that vastly different stages of perceptual decision making rely on common computational principles (Gao et al, Communications Psychology, 2025).