Prior probability cues bias sensory encoding with increasing task exposure

Reviewer #1 (Public Review):

Summary:

Walsh and colleagues investigated how cued probabilistic expectations about future stimuli may influence different stages of decision-making as implemented in the human brain. In their study, participants were provided with cues that could correctly (or incorrectly) cue which visual stimulus would be presented. These cues also predicted the motor action that would likely produce a correct judgment for that trial. In addition, a 'neutral' cue was included that did not predict any particular stimulus. They report that measures of steady-state visual evoked potentials (SSVEPs, proposed to index the magnitude of visual neural activity in favour of the correct response) were smaller when the cue incorrectly predicted the upcoming image, compared to when an accurate cue or a neutral cue was presented. Their primary finding adds to an ongoing debate in the field of decision-making research about how cued expectations may influence how we make decisions.

Strengths:

This study uses a carefully constructed experiment design and decision-making task that allows separation of multiple electroencephalographic (EEG) signals thought to track different stages of decision-making. For example, the steady-state visual evoked potential measures can be cleanly dissociated from more anterior beta-band activity over the motor cortex. They also allow evaluation of how cued expectancy effects may unfold over a number of testing sessions. This is important because the most consistent evidence of expectation-related modulations of electrophysiological measures (using EEG, local field potentials, or single neuron firing rates) is from studies of non-human primates that involved many days of cue-stimulus contingency learning, and there is a lack of similar work using several testing sessions in humans. Although there were several experimental conditions included in the study, careful trial-balancing was conducted to minimise biases due to incidental differences in the number of trials included for analyses across each condition. Performance for each individual was also carefully calibrated to maximise the possibility of identifying subtle changes in task performance by expectation and avoid floor or ceiling effects.

Weaknesses:

Although the experiment and analysis methods are cohesive and well-designed, there are some shortcomings that limit the inferences that can be drawn from the presented findings.

The first relates to the measures of SSVEPs and their relevance for decision-making in the task. In order to eliminate the influence of sporadic pulses of contrast changes that occurred during stimulus presentation, a time window of 680-975 ms post-stimulus onset was used to measure the SSVEPs. The mean response times for the valid and neutral cues were around 850-900 ms for correct responses, and within the same time window for errors in the invalid cue condition. In addition, a large portion of response times in perceptual decision-making tasks are substantially faster than the mean due to right-skewed response time distributions that are typically observed. As it has also been estimated to require 70-100 ms to execute a motor action (e.g., a keypress response) following the commitment to a decision. This raises some concerns about the proportion of trials in which the contrast-dependent visual responses (indexed by the SSVEPs) indexed visual input that was actually used to make the decision in a given trial. Additional analyses of SSVEPs that take the trial-varying pulses into account could be run to determine whether expectations influenced visual responses earlier in the trial. Presenting response time quantile plots may also help to determine the proportions of motor responses (used to report a decision) that occurred during or after the SSVEP measurement window.

In addition, an argument is made for changes in the evidence accumulation rate (called the drift rate) by stimulus expectancy, corresponding to the observed changes in SSVEP measures and differences in the sensory encoding of the stimulus. This inference is limited by the fact that evidence accumulation models (such as the Diffusion Decision Model) were not used to test for drift rate changes as could be determined from the behavioural data (by modelling response time distributions). There appear to be ample numbers of trials per participant to test for drift rate changes in addition to the starting point bias captured in earlier models. Due to the very high number of trials, models could potentially be evaluated for each single participant. This would provide more direct evidence for drift rate changes than the findings based on the SSVEPs, particularly due to the issues with the measurement window relating to the response times as mentioned above.

Reviewer #2 (Public Review):

Summary:

We often have prior expectations about how the sensory world will change, but it remains an open question as to how these expectations are integrated into perceptual decisions. In particular, scientists have debated whether prior knowledge principally changes the decisions we make about the perceptual world, or directly alters our perceptual encoding of incoming sensory evidence.

The authors aimed to shed light on this conundrum by using a novel psychophysical task while measuring EEG signals that have previously been linked to either the sensory encoding or response selection phase of perceptual choice. The results convincingly demonstrate that both features of perceptual decision-making are modulated by prior expectations - but that these biases in neural process emerge over different time courses (i.e., decisional signals are shaped early in learning, but biases in sensory processing are slower to emerge).

Another interesting observation unearthed in the study - though not strictly linked to this perceptual/decisional puzzle - is that neural signatures of focused attention are exaggerated on trials where participants are given neutral (i.e. uninformative) cues. This is consistent with the idea that observers are more attentive to incoming sensory evidence when they cannot rely on their expectations.

In general, I think the study makes a strong contribution to the literature and does an excellent job of separating 'perceiving' from 'responding'. More perhaps could have been done though to separate 'perceiving' and 'responding' from 'deciding' (see below).

Strengths:

The work is executed expertly and focuses cleverly on two features of the EEG signals that can be closely connected to specific loci of the perceptual decision-making process - the SSVEP which connects closely to sensory (visual) encoding, and Mu-Beta lateralisation which connects closely to movement preparation. This is a very appropriate design choice given the authors' research question.

Another advantage of the design is the use of an unusually long training regime (i.e., for humans) - which makes it possible to probe the emergence of different expectation biases in the brain over different timecourses, and in a way that may be more comparable to work with nonhuman animals (who are routinely trained for much longer than humans).

Weaknesses:

In my view, the principal shortcoming of this study is that the experimental task confounds expectations about stimulus identity with expectations about to-be-performed responses. That is, cues in the task don't just tell participants what they will (probably) see, but what they (probably) should do.

In many respects, this feature of the paradigm might seem inevitable, as if specific stimuli are not connected to specific responses, it is not possible to observe motor preparation of this kind (e.g., de Lange, Rahnev, Donner & Lau, 2013 - JoN).

However, the theoretical models that the authors focus on (e.g., drift-diffusion models) are models of decision (i.e., commitment to a proposition about the world) as much as they are models of choice (i.e., commitment to action). Expectation researchers interested in these models are often interested in asking whether predictions influence perceptual processing, perceptual decision, and/or response selection stages (e.g., Feuerriegel, Blom & Hoogendorn, 2021 - Cortex), and other researchers have shown that parameters like drift bias and start point bias can be shifted in paradigms where observers cannot possibly prepare a response (e.g., Thomas, Yon, de Lange & Press, 2020 - Psych Sci).

The present paradigm used by Walsh et al makes it possible to disentangle sensory processing from later decisional processes, but it blurs together the processes of deciding about the stimulus and choosing/initiating the response. This ultimately limits the insights we can draw from this study - as it remains unclear whether rapid changes in motor preparation we see reflect rapid acquisition of new decision criterion or simple cue-action learning. I think this would be important for comprehensively testing the models the authors target - and a good avenue for future work.

Reviewer #3 (Public Review):

Observers make judgements about expected stimuli faster and more accurately. How expectations facilitate such perceptual decisions remains an ongoing area of investigation, however, as expectations may exert their effects in multiple ways. Expectations may directly influence the encoding of sensory signals. Alternatively (or additionally), expectations may influence later stages of decision-making, such as motor preparation, when they bear on the appropriate behavioral response.

In the present study, Walsh and colleagues directly measured the effect of expectations on sensory and motor signals by making clever use of the encephalogram (EEG) recorded from human observers performing a contrast discrimination task. On each trial, a predictive cue indicated which of two superimposed stimuli would likely be higher contrast and, therefore, whether a left or right button press was likely to yield a correct response. Deft design choices allowed the authors to extract both contrast-dependent sensory signals and motor preparation signals from the EEG. The authors provide compelling evidence that, when predictive cues provide information about both a forthcoming stimulus and the appropriate behavioral response, expectation effects are immediately manifest in motor preparation signals and only emerge in sensory signals after extensive training.

Future work should attempt to reconcile these results with related investigations in the field. As the authors note, several groups have reported expectation-induced modulation of sensory signals (using both fMRI and EEG/MEG) on shorter timescales (e.g. just one or two sessions of a few hundred trials, versus the intensive multi-session study reported here). One interesting possibility is that perceptual expectations are not automatic but demand the deployment of feature-based attention, while motor preparation is comparatively less effortful and so dominates when both sources of information are available, as in the present study. This hypothesis is consistent with the authors' thoughtful analysis showing decreased neural signatures of attention over posterior electrodes following predictive cues. Therefore, observing the timescale of sensory effects using the same design and methods (facilitating direct comparison with the present work), but altering task demands slightly such that cues are no longer predictive of the appropriate behavioral response, could be illuminating.