![]() Staying with our first example, Kinchla and Wolfe ( 1979) wrote: What hasn’t changed during this time is our notion of bottom-up and top-down processes. ( 2012), conceptual and methodological advances have substantially increased our knowledge of sensory processing in general and of the visual system in particular. In the three decades between the studies of Kinchla and Wolfe ( 1979) and Melloni et al. Stimulus selection could thus be flexibly adapted by referring to the saliency map most relevant in the context of a given task (Weidner et al., 2009). The authors concluded that multiple saliency maps are present at different levels of processing. Results indicated that primary visual cortex (V1) encodes only bottom-up salience, V2 encodes only top-down control settings, and V4 encodes the interaction between the two. This was manipulated by presenting trials of different degrees of bottom-up salience either in blocks or randomly intermixed. ( 2012) as the influence of our inner goals on stimulus selection. Top-down control is described by Melloni et al. ![]() Thus, the target stimulus could be either a singleton in terms of color and orientation a singleton in terms of orientation only or a singleton in terms of orientation accompanied by a distractor singleton in terms of color. ![]() Bottom-up salience was manipulated by adding color as an additional stimulus dimension. In their visual search task, target and distractor gratings differed in orientation. The authors refer to bottom-up salience as the degree of difference between a stimulus and its neighbors. ( 2012) examined the generation of saliency maps in the visual cortex with functional magnetic resonance imaging (fMRI). Based on this information, the system would then work its way toward more global or more local aspects, as required by the task. Kinchla and Wolfe ( 1979) concluded that visual perception does not proceed strictly bottom-up or top-down, but “middle-out.” They suggested that the visual system uses the information most readily available in the context of, for example, a particular stimulus size and viewing distance. Results indicated that whether a global shape or its constituent elements are processed faster critically depends on stimulus size. Bottom-up processing was defined as “the opposite,” i.e., faster responses if the local stimulus elements, rather than the overall shape, corresponded to a target letter. Upper-case letters were used as stimuli, and top-down processing was assumed to manifest itself in shorter response times (RTs) if the global shape, rather than the local elements, corresponded to a target letter. Following up on earlier work (Reicher, 1969 Navon, 1977), the authors used compound stimuli, where a global shape is made up of smaller, local elements. More than 30 years ago, Kinchla and Wolfe ( 1979) set out to test whether visual processing is organized bottom-up or top-down. We avoid additional terms such as “recurrent” or “reentrant,” because we believe they can be subsumed under “feedback” in most cases.Ī Short History of Bottom-Up and Top-Down Functional activity along these two types of connections is referred to as feedforward and feedback, respectively. Throughout the remainder of this paper, we refer to ascending and descending connections when discussing the anatomy of biological systems or the architecture of artificial ones (Friston, 2005 Clark, in press). We then outline how predictive coding offers a unique perspective for re-defining what is meant by these terms. In this review, we argue that predictive coding provides a powerful conceptual framework that goes beyond the standard dichotomy of “bottom-up” and “top-down.” We first provide an overview of previous attempts at defining bottom-up and top-down processes, and we highlight their problems and limitations. In accordance with terminology commonly used in research on perception, it has thus been suggested that predictive signaling reflects top-down processes, whereas prediction-error signaling constitutes bottom-up processing (Friston, 2009 Alink et al., 2010 Hesselmann et al., 2010). Evidence for this assumption has been collected at different levels of neural processing (Rao and Ballard, 1999 Hosoya et al., 2005 Muckli et al., 2005 Summerfield et al., 2006 Alink et al., 2010), which suggests that predictive-coding operates across a wide range of spatial and temporal scales.Ī second important assumption in predictive coding is that predictions are transferred from hierarchically higher levels of processing to lower ones, whereas signals traveling in the opposite direction encode prediction errors (Rao and Ballard, 1999 Serences, 2008 Friston, 2009 Grossberg, 2009). A central assumption in predictive-coding theories is that activity in the nervous system reflects a process of matching internally generated predictions to external stimulation (Heekeren et al., 2008 Bar, 2009).
0 Comments
Leave a Reply. |