Resumen de Attentional influences on learning under unpredicted situations
Among the wide variety of stimuli that we process everyday just a small portion enters our conscious mind, and even a smaller fraction is efficiently encoded into memory and can be later remembered. How does our system select which information will be made available to consciousness? What are the characteristics of a given event that would render it relevant enough to be encoded into memory? This doctoral dissertation aims at offering insight into these and other related questions by framing the broader topic within a predictive coding approach. Assuming that cognition works by default by anticipating the most likely input and setting up the system to respond adaptively to this predicted input, this work will present different ways to address the overall question of how introducing conflict or breaking up expectations modulate different stages of cognition.
These hypotheses were tested in four experimental series, presented in twochapters, thus groupingthe experimental series into two major categories, based on a methodological distinction: in the first chapter weuseverbal materials to explore the long-term consequences of spatial expectation mismatches; in the second chapter weuse ecological visual materials to study both online and offline consequences of semantic conflict.
The first chapter aimsto ascertain whether there is a core learning mechanism triggered by mere expectation mismatch. Several neural and computational models predict that breaking an expectationshould trigger a learning signal that enhances encoding of the episode as a wayto improve future predictionsof thisformerly unexpected event (Henson & Gagnepain, 2010; Verguts & Notebaert, 2009). This expectation mismatch, or prediction error, as it is known within the predictive coding framework, is thought to be domain general and therefore to result from any expectation violation. Whereas it is possible to find studies that showthese memory improvements afterdifferent unfulfilled expectations, they often include a number of confounding factors (Krebs, Boehler, De Belder, & Egner, 2015; Rosner, D’Angelo, MacLellan, & Milliken, 2015; Rosner, Davis, & Milliken, 2015). In this experimental series we aimed at solving this confound by using the most basic and aseptic paradigm capable of inducing an expectation: a Posner cueing paradigm. Across seven experiments we thoroughly manipulated many commonly studied variables to show that mere expectation mismatch is not able to enhance memory as attested by a surprise recognition test. This finding is important to shapecurrent models of prediction error-induced learning, since it implies that either a more cautious definition of prediction error needs to be established, or a revaluation of the major assumptions of these modelsis required.
The second chapter makes use of ecological scenes to explore expectation mismatches at a semantic level. Scene processing is known to proceed gradually from the gist of the scene, that is grasped very early (Castelhano & Henderson, 2008; Oliva & Torralba, 2006; Sampanes, Tseng, & Bridgeman, 2008), to the identification of individual objects, that may take considerably more time. Under this framework, we assumed that individual object processing can also be understood in terms of anticipations of the most likely candidates based on the available semantic information (Rao & Ballard, 1999; Summerfield et al., 2006). Therefore, the experiments integrated in this chapterinclude objects thatare semantically incongruent with their surrounding background, to induce a form of semantic prediction error. We used a change detection task in which we manipulated the semantic congruity between the target objects and the background scenes in which they were embedded. In the three experimental series included within this second chapterwe used this general method to explore how semantic mismatches impact cognitive processing at different stages ranging from pure detection, through identification, and to eventual remembering.
The first experimental series of this second chapter presents a counterintuitive effect of semantic mismatch on remembering. As stated above, current models on conflict-induced learning predict that better memory should be obtained for context-incongruent objects (Henson & Gagnepain, 2010; Verguts & Notebaert, 2009). However, our results show that the opposite is true. Those targets that are semantically congruent with their surrounding scene were remembered better than those that did not. This unexpected result is in line with the findings obtained in the first chapter and thus supports the need for reevaluating the idea that expectation mismatches are always associated to an increase in memory or learning. Moreover, the results indicate that, at least under certain conditions, structured and coherent items could be better encoded into memory than are either unexpected or incongruent ones.
The second experimental series of this chapter further explores this puzzling result by analyzing three behavioral indexes that serve as proxies for three stages in the stream of processing of visual information. Namely, it has been found that the same factor of semantic congruity produces opposite effects at different processing stages, giving place to a detection cost, an identification benefit and a recognition benefit. In this study, we attempt to characterize the mechanisms underlying each of these three indexes, as well as theeye-movements’ patterns associated with them. Our results of participants performance showed that i) the detection cost, but not the identification benefitnor the recognition benefit, is contingent onthe presence of object competition and that shorter and earlier fixations on incongruent targets are responsible for the detection cost, ii) that the identification benefit is not associated to the most commonmeasures of eye movement and iii) although better memory in general is observed for better identified objects,the recognition benefit is not dependent on the level of object identification. To sum up, the results from the third experimental series further support that semantic characteristics of the stimuli affect processingfrom the earliest stages of conscious processing to eventual remembering.
Finally, the last experimental series of the second chapter, aimed at analyzing these semantic influences on object processing at a neural level. We used Transcranial Magnetic Stimulation (TMS) to suppress activity in a brain region that has been related with attentional reorienting, namely, the right temporo-parietal junction (TPJ), and assessed whether this manipulation would affect the detection cost orthe identification benefit. Although activity in the entire right parietal lobe has been traditionally implicated on attentional tasks (Geng & Mangun, 2011; Serences et al., 2005), the right TPJ is specifically thought to be responsible for attentional re-orienting in space (Bourgeois, Chica, Valero-Cabré, & Bartolomeo, 2013; Corbetta, Kincade, Ollinger, McAvoy, & Shulman, 2000). We hypothesized that, given that the detection cost, but not the identification benefit, seems to arise as a consequence of efficient attentional attraction toward the target region, suppressing activity in right TPJ would reduce, or even eliminate, the detection cost, while leaving unaffected the identification benefit. Our results show that repetitive TMS over right TPJ effectively reduced the detection cost compared to the stimulation of a control region (vertex). Importantly however, the identification was also reduced by right TPJ stimulation. This double and partially unexpected finding implies that,although a clear dissociation between detection and identification processes can be behaviorally established, both processes most likely interact with each other by means of some shared underlying neural substrate. Moreover, the attentional role traditionally assignedto right TPJ might need reevaluation. Indeed, this result is in line with other recent research attributingto this region the role of contextual updating rather than that of re-orienting attention (Doricchi, Macci, Silvetti, & Macaluso, 2010; Geng & Vossel, 2013). In our data set,impairing of efficient contextual updating could account for the reduction of both the detection cost and the identification benefit.
In sum, the present thesis reveals the need for rethinking those models that characterize learning as generally bolstered by any kind of conflict or prediction error. It also presents an interesting and promising perspective on the study of semantic influences at different levels across the stream of processing, from very early stages of object detection, to identification and memory. Finally, the present work supports a new conceptualization of the role of right TPJ on attentional processes (Doricchi, Macci, Silvetti, & Macaluso, 2010; Geng & Vossel, 2013) and opens new questions regarding the brain areas implicated in congruity-mediated recognition processes.
