Visual detection of time-varying signals: Opposing biases and their timescales

Human visual perception is a complex, dynamic and fluctuating process. In addition to the incoming visual stimulus, it is affected by many other factors including temporal context, both external and internal to the observer. In this study we investigate the dynamic properties of psychophysical responses to a continuous stream of visual near-threshold detection tasks. We manipulate the incoming signals to have temporal structures with various characteristic timescales. Responses of human observers to these signals are analyzed using tools that highlight their dynamical features as well. Our experiments show two opposing biases that shape perceptual decision making simultaneously: positive recency, biasing towards repeated response; and adaptation, entailing an increased probability of changed response. While both these effects have been reported in previous work, our results shed new light on the timescales involved in these effects, and on their interplay with varying inputs. We find that positive recency is a short-term bias, inversely correlated with response time, suggesting it can be compensated by afterthought. Adaptation, in contrast, reflects trends over longer times possibly including multiple previous trials. Our entire dataset, which includes different input signal temporal structures, is consistent with a simple model with the two biases characterized by a fixed parameter set. These results suggest that perceptual biases are inherent features which are not flexible to tune to input signals.


Fixed threshols in all stimulus regimes
Comparing thresholds of psychometric curves fit to individual observers, we see no change in threshold across the three stimulus regimes. This is true also on average over all observers. Here we used for the instantaneous model psychometric curves with slopes that vary across input signal type. The results reported in the main text were obtained using the average curve over the entire experiment; no significant change was found.

Figure C: POA is not sensitive to differences in psychometric curves among stimuli.
In this analysis the curves slope for the instantaneous observer are different for each input mode. Specifically, we used k w = 29.3, k p = 32.8, k b = 36.5 for White, Pink and Brown respectively, which are the average slopes over observers in each regime. The results are very similar to those in the main text where k = 30 in all cases

Models of Separate Biases
Two partial models were tested separately, with only one of the two biases incorporated in each. One model accounts for the positive recency bias only, and the other for the long-term adaptation only. These two effects are combined in the model that was presented in the main text in section and its performance is presented in the figures below. The hysteresis with respect to direction of input trends is missing the negative phase that dominates in the long τ values in the data. (d) POA averaged in each stimulus regime was used to calibrate the model with the correct strength of positive recency. The difference between the POA in this partial model to the actual data is therefore zero within accruacy of the measurements. (d) The difference between this partial model and the data POA is increasing, reflecting higher alternation rates than as correlation time increases. This is in qualitative disagreement with the data.

Other models tested
Bias modulations by response only This model is similar to the one presented in the main text, only with the adaptation bias regulated by the history of responses rather than of inputs. The output y is filtered with a time constant τ to give y i , which replaced x i in the adaptation variable A. The performance the two models is almost identical, however we find this configuration of feedbacks to be less plausible physiologically.

Bias modulations of output only
In this model the two biases are introduced in the post-sensory stage of processing. Sensory processing is instantaneous and independent on history. It is modeled by constant sigmoidal relations between the momentary input level and the probability of the coin flip. The decision itself, on the other hand, encapsulates all history dependencies. The performance of this model was inferior to the one presented in the main text in reproducing the experimental results, and moreover parameter values needed to be tuned and the results were less robust.

Spatial Effects
Here we tested the influence of spatial proximity on the response. Different distances in pixels were used to define spot as "close"/"far" to the previous one. A reduction of threshold was found for spatially close-by spots, (example in Figure Ha(a)), but only when spot location was very close, less than a half the spot size (distance<=30 pixels between centers) [Druker, 2010]. The effect reduced rapidly with distance ( Figure   H(b)), diminishing completely when no overlap existed between consecutive stimuli (distance>60 pixels). Trials where the distance between consecutive spots was <=30 pixels constitute only 15% of all trials ( Figure H(c)). Therefore, this small fraction alone cannot account for the global effect of adaptation. Figure H: Spatial Effects (a) Example of threshold reduction by consecutive overlapping spots, with less than 30 pixels distance between centers (triangles), relative to the rest of the trials (circles). Insert: for all observers the thresholds psychometric curves of "far" stimuli were subtracted from those of the "close" ones. The differences were overall small and negative. (b) This spatial effect was tested with various threshold distances. The differences were detected only in the very close distances, when the consecutive spots were partially overlapping (distance<=30 pixels).(c) The fraction of "close" distances out of all trials, as a function of defined pixel overlap. (d) The detection probability of very short distances between consecutive trials is slightly higher than these of the long distances. The individual differences between the DP are plotted for of short of distances <30 pixels, i.e. less than half overlap between the consecutive trials. (e) and (f) no consistent effects on slope and probability of alternation were found.