Measures of repetition suppression in the Fusiform Face Area are inflated by co-occurring effects of statistically learned visual associations

Repeated presentation of a stimulus leads to reductions in measures of neural responses. This phenomenon, termed repetition suppression (RS), has recently been conceptualized using models based on predictive coding, which describe RS as due to expectations that are weighted toward recently-seen stimuli. To evaluate these models, researchers have manipulated the likelihood of stimulus repetition within experiments. They have reported findings that are inconsistent across hemodynamic and electrophysiological measures, and difficult to interpret as clear support or refutation of predictive coding models. We instead investigated a different type of expectation effect that is apparent in stimulus repetition experiments: the difference in one’s ability to predict the identity of repeated, compared to unrepeated, stimuli. In previous experiments that presented pairs of repeated or alternating images, once participants had seen the first stimulus image in a pair, they could form specific expectations about the repeated stimulus image. However they could not form such expectations for the alternating image, which was often randomly chosen from a large stimulus set. To assess the contribution of stimulus predictability effects to previously observed RS, we measured BOLD signals while presenting pairs of repeated and alternating faces. This was done in contexts whereby stimuli in alternating trials were either i.) predictable through statistically learned associations between pairs of stimuli or ii.) chosen randomly and therefore unpredictable. We found that RS in the right FFA was much larger in trials with unpredictable compared to predictable alternating faces. This was primarily due to unpredictable alternating stimuli evoking larger BOLD signals than predictable alternating stimuli. We show that imbalances in stimulus predictability across repeated and alternating trials can greatly inflate measures of RS, or even mimic RS effects. Our findings also indicate that stimulus-specific expectations, as described by predictive coding models, may account for a sizeable portion of observed RS effects.

by contextual factors, such as the probability of stimulus repetition. They presented pairs 82 of faces in each trial and reported that BOLD signal differences between repeated and 83 unrepeated stimuli (i.e., repetition effects) were larger in blocks with high (75%), 84 compared to blocks with low (25%) repetition probability. This interaction involving 85 repetition probability was replicated several times using faces (for a review see Grotheer et 86 al., 2014), and also for other stimulus categories such as letters (Grotheer & Kovács, 2014) 87 and other non--face objects (Kronbichler et al., 2018;Mayrhauser et al., 2014). Notably, this 88 interaction has mostly been reported in studies using fMRI; when using  stimulus was repeatedly paired with the first stimulus during a prior training session.

160
They observed large differences in the magnitude of repetition effects, apparently due to 161 reductions in BOLD signals for predictable compared to unpredictable alternating faces.

162
Further evidence for predictability effects came from a recent EEG study (Feuerriegel et 163 al., 2018a), who used a similar blocked design with predictable and unpredictable 164 alternating faces. In the so--called "AB" blocks in that experiment the second stimulus in 165 each trial could either be the same image as the first (repetition trials), or a specific same--166 sex face (predictable alternation trials). In the "AX" blocks, however, the second stimulus 167 could either be a repetition of the first one, or a same--gender face, selected randomly from 168 a set of 23 stimuli (unpredictable alternation trials). Differences in event--related potential 169 (ERP) repetition effect magnitudes across AB and AX blocks were found during multiple 170 time windows post stimulus onset. Importantly, these differences in observed repetition 171 effects were due to differences in ERP responses to alternating stimuli across block types, 172 and no differences across AB and AX blocks were found for repeating stimuli. 173 Critically, this study did not equate the relative novelty of AB and AX alternating  In each trial (Fig.  1A) an adapter (S1) and test stimulus (S2) were each presented for 229 250 ms, separated by an inter--stimulus interval (ISI) of 400--600 ms (randomised across 230 trials). The image size of S2 was 20% smaller than that of S1 to avoid low--level adaptation 231 processes. Trials were separated by an inter--trial interval (ITI): for the training sessions 232 the ITI was 1800, 2000 or 2200 ms, randomly distributed across trials, and for the fMRI 233 sessions it was 6, 8 or 10 seconds.

234
In each trial, S1 and S2 could either be identical (repetition trials; Rep) or depicting 235 different identities (alternation trials; Alt). These trial types were presented in two 236 different contexts (Fig.  1B), labelled as "AB" and "AX". In the AB context the S2 face could 237 either be a repetition of the S1 face (Rep trials), or a specific face identity that had 238 previously been repeatedly paired and associated with the S1 identity during the training 239 sessions (Alt trials). In these Alt trials of the AB context, each S1 face identity was 240 consistently paired with one of the five other face identities that were allocated to the AB 241 context. Each S1 identity in the AB stimulus set was paired with a different S2 face 242 identity, ensuring that each face image would be presented an equal number of times 243 throughout the experiment. In other words, once the participant has seen a given S1 face 244 "A", they could form expectations regarding the S2 to be a repetition of face "A" or a 245 different, specific identity "B". In the AX context S2 could either be the repetition of the S1 246 image, or a different identity, pseudo--randomly selected from the set of 5 other face 247 identities. Therefore, in the AX context, there were no consistent pairings between S1 and  (2), --21 (1); p < 0.05 FWE) and included this ROI in a separate analysis.

304
Magnetic Resonance Images were acquired using a 3--Tesla magnetic resonance

355
Peak BOLD signal amplitudes by participant and condition are displayed in Figure  2A. We  361 We also found an interaction between context and trial type in the rFFA (F (1,17) =  using an AB--type design, with associated stimuli and an S1--S2 matching task.

463
In addition, we note that the RS effects in our study may not be strictly localized to    513 We have shown that, in immediate repetition designs, an observer's capacity to predict 514 the image of repeated compared to unrepeated stimuli has a substantial effect on the 515 observed magnitude of RS. While this does not necessarily mean that RS is best accounted 516 for by predictive coding models, it does indicate that measures of repetition effects have 517 likely been inflated due to this confound in a very large number of previous studies, 518 including those run within our own labs. We also highlight stimulus predictability as an 519 important, yet commonly overlooked, factor to consider when investigating the hierarchy 520 of expectation effects implemented within the visual system.