1.1 Incorporating learning and memory into perceptual development

Since infants are engaging in sophisticated types of learning and memory during the same developmental period when they are experiencing rapid changes in their visual development (i.e., before 12 months of age), it is essential to determine whether learning and memory is influencing experience-based changes in vision, and perception more broadly, early in life. With this in mind, the goal of this article is to present two opposing types of theories or models of perceptual development. One allows the impact of these learning and memory abilities directly on perception starting early in life and the other type does not. The key difference between these two types of models is the availability of neural connections from cortices supporting learning and memory (frontal cortex, MTL, striatum) and perceptual systems. These neural connections are feedback connections: For example, the striatum cannot influence the feed-forward input into the visual system; instead, it must influence the visual system through feedback connections. For this reason, we consider these two contrasting models as exclusively feed-forward/bottom-up and feedback/top-down.

Feed-forward/bottom-up models of perceptual development have been dominant historically and are consistent with the view that perceptual learning and development is separate from the sophisticated learning and memory abilities that infants possess. For example, Aslin and Smith (1988) argue that early vision is constrained by limitations in the sensitivity of the perception system (e.g., reductions in visual spatial acuity) and, when these low-level constraints are relieved later in the first year of life, this allows developmental changes in the representation of perceptual input (e.g., objects or forms) that emerge later in the first postnatal year (e.g., 6–12 months). Then, heading into childhood, one sees the emergence of more cognitive abilities (e.g., task-based attention) to support yet further developments in perception. Thus, low-level perception develops, then higher-level perception develops, then more cognitive factors impact perception after these initial stages are over. They do not consider the role of memory or familiarity on visual development. Since low-level perception, high-level perception, and cognition are considered along a progression from earlier to later in the cognitive or cortical hierarchy, this is an exclusively feed-forward/bottom-up model of perceptual development. Despite their historical importance and early usefulness, in recent years, the field has been questioning these exclusively feed-forward/bottom-up models of perceptual development.

By contrast, feedback/top-down models of perceptual development will allow more influence of not only learning and memory, but other cognitive factors such as language, on perceptual development. If these abilities are found to influence perception, perceptual developmental cannot be exclusively feed-forward and bottom-up but must incorporate feedback neural connections to allow for the top-down influence of perception starting early in life.

There is not sufficient evidence to argue for one or the other of these models at this time. Thus, the goal of this article is to consider these types of models as clearly as possible in order to facilitate future work. To this end, I will outline the different commitments of these models, highlight some of the benefits and limitations of each and present some of the evidence. The ultimate motivation for this article is not to argue for or against a particular view but to move the field forward in the investigation of the mechanisms supporting perceptual development.

2.1 Models of visual development are predominantly and historically bottom-up

While the field of perceptual development has not historically created explicit models, the models classically employed have been bottom-up. Eleanor Gibson (e.g., 1969) excluded any cognitive influences on perceptual development and, even more extremely, also rejected constructivist approaches such that representations, arguably part of higher-level vision, were not included. Instead, the infant or learner discovered invariants in their sensory input that occur over time and space that allow them to see the world directly. Gibsonian perceptual development has a largely flat hierarchy where perception development occurs in isolation from other cognitive abilities.

Aslin and Smith's (1988) comprehensive review of perceptual development explicitly lays out a very similar hierarchy to what is presented here with sensory primitives (corresponding to low-level perception), perceptual representations, and cognitive and language influences (see also summary of this model above). When considering visual development, the authors explicitly state that the hierarchy is bottom-up where the first phase of development is constrained to sensory primitives and, when sensory primitives are sufficiently developed, that permits the subsequent development of perceptual representations. Once there has been adequate development of perceptual representations, the authors argue that in early childhood there can be other (cognitive) influences on vision.

Turning now to more contemporary models, a major shift in the field of perceptual development in the 21st century was the discovery of perceptual narrowing as a visual and, arguably, domain-general, perceptual developmental phenomenon. Very young infants have the capacity to equally (but poorly) perceive many types of faces (e.g., human and monkey faces, faces of all races) and many types of speech sounds (i.e., those of many different languages). Infants differentially develop perceptual abilities for the faces and speech sounds that are present in their environment compared to ones that are not present in their environment (faces: Kelly et al., 2007; Pascalis et al., 2005; speech: Werker & Tees, 1984). This phenomenon of perceptual narrowing presents a compelling case that an infant's experience shapes their perceptual development in a way that supports the emergence of more sophisticated perceptual abilities and has been the focus of the field of perceptual development for most of the 21st century.

Theorizing about how perceptual narrowing occurs has focused largely on the debate about whether it arises from sensory experience or is mediated by biologically-determined critical periods. This particular debate is beyond the scope of this review. But, focusing on the models of how experience supports perceptual narrowing, both early and more recent reviews have argued that it is mere sensory experience that drives this phenomenon as opposed to the involvement of learning and memory or other more cognitive systems. Scott, Pascalis, and Nelson (2007) note that the development of specialization “corresponds to improved perceptual discrimination for stimuli predominant in the environment relative to declining perceptual discrimination for stimuli not present in the environment” (p. 200) and point to synaptic pruning and Hebbian learning as potential neural mechanisms supporting perceptual narrowing. Through these mechanisms, “as perceptual experience with human faces increases, the strength of the neural circuit responding to human faces is strengthened” (p. 201). A more recent review suggests a similar mechanism: Maurer and Werker (2014, p. 171) suggest that it is the “simple accrual of experience” that results in perceptual narrowing. Both of these models point to differences in an infant's sensory experience as the cause of perceptual narrowing (e.g., Caucasian infants predominantly see Caucasian faces). These explanations speak to the view that perceptual development only requires the interaction of sensory input and a developing perceptual system that are necessary to result in perceptual development with development occurring at the level of relevant representation (i.e., face or speech) rather than involving the interaction from cognitive systems such as learning and memory located beyond the perceptual systems that may be activated by this differential experience.

2.2 Commitments of bottom-up models

The explicit architecture of the bottom-up models results in particular commitments for how perceptual development should proceed. These commitments can be operationalized as testable hypotheses and will be considered in relation to the recent empirical findings. Table 1 summarizes these commitments of exclusively feed-forward/bottom-up models.

A key commitment of an exclusively feed-forward/bottom-up model is that perceptual development is entirely stimulus-driven. In other words, what sensory input an infant receives drives what develops. This constraint results in the strong prediction that if infants receive the same sensory input that they will have the same development outcomes; For example, if two infants see the same distribution of faces, they will have the same face discrimination abilities.

Feed-forward/bottom-up models would result in development that is, to some extent, sequential from lower- to higher-levels of perception. Drawing from Aslin and Smith (1988), sensory primitives must be adequately developed before development at the higher-representational level is possible and which then facilitate cognitive abilities later on development. Indeed, sequential development is a common feature in experience-based models of language development. Since there is no strong evidence of sequential development (e.g., speech perception develops before lexical representations), the requirement of sequential development has been recognized as a key weakness in these models (see speech sounds and word learning or lexical development: Saffran & Thiessen, 2007; Swingley, 2008).

Bottom-up models predict that learning and development will be local and arise slowly. Bottom-up models invoke mechanisms like Hebbian learning and pruning (e.g., Scott et al., 2007) which can strengthen representations of stimuli that an infant experiences. While it is not worked out how these mechanisms would give rise to perceptual learning, these mechanisms result in structural changes in the cortex through modifying the strength and/or distribution of synapses in the cortex. Since they are working at the level of individual synapses, they will arise from highly local learning that is specific to the level of the relevant representations (i.e., Hebbian learning of faces will occur in regions of the cortex that are selective for faces). Moreover, since Hebbian learning and pruning are mechanisms that alter the structure of the cortex, these types of changes will arise over long periods of time and a great deal of experience. For example, there is an important developmental period of pruning that occurs over many years (and even decades) starting early in life (Webb, Monk, & Nelson, 2001). Recent advances in imaging technologies (e.g., two photon microscopy) have allowed researchers the ability to see how synapses change in response to experience in animal models. While this work has revealed that synapses are very dynamic and change their structure as a result of experiences over days (as opposed to years), these synaptic changes are produced through extreme variations in sensory input and usual complete deprivation of input (e.g., from a limb: Zuo, Yang, Kwon, & Gan, 2005; from one or both eyes; see Zhou, Lai, & Gan, 2017). Given that the differences in sensory experience arising from seeing only Caucasian faces as opposed to seeing a wider variety of faces are very different from complete visual deprivation, it is unlikely that synaptic changes arising from face experience would modify synapses in order to affect experience even on the scale of days. Thus, the mechanisms proposed by bottom-up models of perceptual development will result in highly local and slow changes in perception.

The information flow of feed-forward/bottom-up models is unidirectional from the lowest levels of the visual hierarchy moving upward toward cognitive systems without any feedback or top-down connections. Without these feedback connections, perceptual development is passive: It occurs without any cognitive engagement from the infant, and without the interaction of higher-level to lower-level perception. In an extreme expression of this idea, perceptual experience could support perceptual development without any region in the brain beyond the perceptual cortices.

Finally, bottom-up models will result in changes in perception that are quite robust but also inflexible. Since structural changes in the cortex are invoked (e.g., Hebbian learning and pruning), sensory experience will result in perceptual changes that have resulted from structural changes in the cortex and, therefore, will affect perception regardless of task context. For example, if perceptual narrowing results in a structural strengthening of representations for the faces that an infant has seen, the perceptual benefit of this experience should be present regardless of the context in which that face occurs.

4.1 Commitments of a top-down model of perceptual development

Top-down models of perceptual development make a series of different commitments from an exclusively bottom-up model (Table 1). First, the flow of information and systems available to influence perceptual development are much less constrained. While bottom-up changes and feed-forward information are still available in a top-down model, there is an additional feedback from higher- to lower-levels of the visual system and from the myriad of cognitive systems (learning and memory, reward, attention) or systems that would require feedback connections to influence cortical vision systems (e.g., auditory). Thus, a top-down model allows for many more interactions between segments of this hierarchy through the introduction of feedback connections (see Gilbert & Li, 2013 for another paper that equates feedback connections with top-down modulation).

Since a top-down model is much more interactive, the changes and systems that give rise to changes in perception can be more distributed. In other words, while some aspects of perceptual development can arise from highly local changes (e.g., changes in face perception arising from changes locally to the fusiform), they may also depend on regions or systems located higher in the hierarchy (e.g., changes in face perception may arise from the engagement of attentions systems likely located in the fronto-parietal regions of the brain, see Markant et al., 2016; Markant & Scott, 2018). The dependence of perceptual changes on these other regions is not possible in an exclusively feed-forward/bottom-up model of perceptual development but is possible in a feedback/top-down model. A number of neuroimaging studies in adults have now pointed to the involvement of non-visual cortices in experience-based changes in perception (e.g., involvement of the medial temporal lobe in visual changes: Emberson & Amso, 2012; Stokes, Atherton, Patai, & Nobre, 2012; Turk-Browne, Scholl, Johnson, & Chun, 2010). These studies suggest that while perceptual changes occur as a result of experience, these changes occur in concert with regions located beyond the visual cortex.

Because of these interactions and the influence of feedback neural connections, the same sensory input can result in different developmental changes in perception. While an exclusively feed-forward/bottom-up model is strongly constrained by the sensory input an infant receives, a feedback/top-down model allows that the same sensory input can result in differences in perceptual outcomes if and only if there are differences in the structural/cognitive/task context in which the sensory input is received. Thus, two infants who receive the same sensory experience with faces could have differences in their perceptual development if the context or task of this exposure were different. As outlined above in Section 2, this impact of task has already been demonstrated by Scott and Monesson (2009) where infants received the same experience with monkey faces but differed in the context in which this experience was given. This difference in task (i.e., individuation vs. non-individuation or no labels) corresponded to differences in infant perception.

In addition to the slow and gradual tuning that feed-forward/bottom-up models allow, a top-down model allows that experience can shape perception through rapid and flexible changes. Top-down changes in perception do not have to arise from structural changes in the brain (e.g., pruning and Hebbian learning) that alter the representations or sensitivity to particular representations. Instead, these rapid and flexible changes arise from the modulation of perceptual cortices through feedback connections from other systems. To refer to a canonical example, task-driven or endogenous attention has been found in adults to modulate sensory responses in the visual system rapidly (i.e., it does not require months of training or exposure) and flexibly depending on task (e.g., Luck, Fan, & Hillyard, 1993; McMains & Kastner, 2011). While task-driven or endogenous attention is a canonical example, flexible changes in vision are in principle able to be gained from many different modulatory systems including learning and memory (e.g., Peterson, De Gelder, Rapcsak, & Gerhardstein, 2000), reward (e.g., Serences, 2008), and prediction or expectation (e.g., Kok, Rahnev, Jehee, Lau, & De Lange, 2012; Summerfield & de Lange, 2014). All of these changes in the visual system can be seen after a very short period of exposure (i.e., are rapid) and are, in some cases, dependent on context.

Even though top-down processes can give rise to rapid changes, mediated through top-down processes, they can still result in lasting changes in perception. Folstein, Palmeri, Van Gulick, and Gauthier (2015) argue that perceptual category learning employs selective attention to expand the perception of certain feature dimensions and these changes both arise rapidly and persist over weeks after only a short period of training.

4.2 Evidence for top-down perceptual development in infancy

In Section 2, several behavioral studies are reviewed that indicate that perceptual development is not only driven by the sensory input that an infant receives. Beyond these challenges to an exclusively feed-forward/bottom-up model, recent behavioral and neuroimaging studies have established that young infants already have the capacity to employ top-down modulation to change activity in their visual system and that disruptions of this ability are associated with poor developmental outcomes. Moreover, there is some initial evidence from clinical studies that the striatum is involved in visual development.

Emberson, Richards, and Aslin (2015) examined responses in the visual systems of 6-month olds when they predicted a visual stimulus. This study used functional near-infrared spectroscopy (fNIRS), a light-based method of recording the hemodynamic response, the same physiological substrate as fMRI, in the surface of the infant cortex (Aslin et al., 2015). Consistent with the view that top-down mechanisms are available early in development, infants exhibited robust visual cortex responses when a visual stimulus was unexpectedly omitted. In other words, visual expectations can provide feedback to drive perceptual cortex responses in infancy as young as 6 months of age.

Importantly, deficits in top-down modulation are associated with poor developmental outcomes. Premature birth has been found to put infants at-risk for language delays, motor deficits and learning disabilities. Emberson, Boldin, Riccio, Guillet, and Aslin (2017) found that, at 6 months, infant born prematurely (< 33 weeks gestation) exhibited deficits in neural signals associated with the prediction of visual stimuli. Finding these deficits in top-down processing very early in development suggests that deficits in prediction give rise to developmental difficulties.

Moreover, an isolated study has suggested that the striatum is more involved in visual system development. Mercuri et al. (1997) investigated 37 infants with perinatal brain lesions to determine the locus of lesion that results in visual impairment as measured using behavioral and electrophysiological measurements. Consistent with the view that the young brain is highly plastic, even infants with bilateral occipital lesions had normal visual function. However, infants with basal ganglia lesions (either associated with occipital lesions or isolated) had abnormal visual function even when assessing vision using non-motor tasks (e.g., VEP, p. F111).

While this study bears follow-up, the finding that the striatum, a region not typically associated with visual function, should be considered in relation to the broader concept of developmental diaschisis. Coined in Wang, Kloth, and Badura (2014), developmental diaschisis is a concept where disruptions in a particular brain area during development, such as the basal ganglia in this case, can affect the development of disparate regions of the brain but not to function in the mature brain. Using this concept, Wang et al. (2014) propose that the cerebellum, a region typically associated with motor function in the mature brain, is involved in the development of non-motor cortical circuits. As part of evidence to support this, they review compelling evidence that abnormal cerebellar development increases an individual's risk for developing autism spectrum disorder (ASD) despite cerebellar dysfunction in adulthood having no such association. Returning to the striatum, Fuccillo (2016) also proposes the striatum to be involved in ASD through developmental diaschisis. Mercuri et al. (1997) suggest that while the basal ganglia is not important for visual function later in life, it is important for early visual development. This type of observation or the existence of developmental diaschisis could not be accommodated in an exclusively bottom-up model of perceptual development.

Together, these findings provide some initial evidence that feedback connections are available to modulate the visual system starting early in infancy (Emberson et al., 2015), that impairment in this top-down modulation is associated with infants at-risk for developmental delays (Emberson, Boldin, et al., 2017) and that regions located outside the visual system may be more essential for an infant's visual function than the occipital cortex (Mercuri et al., 1997).

5.1 Scott, Markant and colleagues: Top-down attention and face perception

The most related proposal to the current one has been put forward by Scott, Markant and colleagues in two papers: Hadley, Rost, Fava, and Scott (2014) and Markant and Scott (2018). These proposals are the first to consider whether top-down attention can play a role in supporting perceptual narrowing of faces in infancy. In particular, Markant and Scott (2018) present an exciting, ground-breaking and concrete model of how early bottom-up attentional biases can interact with latter developing top-down attentional biases to give rise to the kind of perceptual selectivity that is seen at the end of the first postnatal year (e.g., 10–12 months). While there are many points of convergence between the current proposal and the model outlined by Markant and Scott, we highlight two major differences between these proposals.

Scott, Markant and colleagues' model is highly specific to face perception and attentional development. It would not be trivial to simply broaden the Markant and Scott's (2018) proposal to other aspects of vision for two reasons: (1) Their model depends on particular bottom-up attentional biases to faces early in life; (2) They propose that the changes in face perception after 6 months are specifically tied to developmental changes in endogenous attention specific to face perception.

Second, the proposals have different definitions of top-down and the mechanisms of action for top-down attention may be distinct between the proposals. The Scott, Markant and colleagues’ proposals rely heavily, though not exclusively, on findings where infants' eye-movements to faces are altered over development as evidence that top-down attention is shaping face perception. While it is the case that sampling different aspects of the visual input results in differences in sensory input, it is not clear whether these subtle, if systematic, changes in eye movements meaningfully change perception (see below) and whether these changes in eye movements arise through use of feedback connections from the anterior attentional system to the visual system. Thus, while Scott, Markant and colleagues' proposals use the term top-down, selective or endogenous attention, it may be that none of these changes rely on feedback neural connections. Indeed, as outlined above, the current proposal uses a concept of top-down that is different from the concepts of top-down and bottom-up attention. Thus, these proposals employ different definitions of top-down and, as such, may be mechanistically independent. It is also possible that there is a great deal of convergence here but additional evidence and clarification about the role of feedback connections would need to be more carefully evaluated in their proposal.

5.2 Johnson, Westermann, and Mareschal: Interactive specialization and connectionist models of object categorization

This proposal owes large debt to the theoretical view of Interactive Specialization (e.g., Johnson, 2011). However, this important theoretical view has not yet considered feedback or top-down influences: Johnson (2011) highlights the question of “how the specialization of early sensory areas is shaped by top-down feedback” (p. 19) as an important Future Direction and notes several relevant adult findings but no developmental evidence for top-down feedback.

While the theoretical work in this literature has not yet been extended in this direction, one relevant model has been published which suggests that the hippocampus could affect perceptual representations of object categories in development through feedback connections (Westermann & Mareschal, 2012, 2014). In brief, this model has a module characterized by a slow learning rate (cortex) and a fast learning rate (hippocampus) with mutual connections between them. When the fast learning rate module is present, there is more separation between the representations in the slow learning module (cortex) than if it is absent. This model is convergent with the current proposal as connections from the hippocampus to the visual cortex would be feedback (i.e., descending the cortical hierarchy). Moreover, the authors explicitly argue that these changes occur “beyond attending to stimuli differently” (p. 7) and integrate changes in perception.

The differences between that proposal and the current one are threefold. First, there is a clear difference in scope: Westermann and Mareschal (2012, 2014) focus exclusively on the hippocampus and object categorization. Second, it is not clear that the changes found in the cortical module are perceptual. They are certainly representational which is exciting but there are no non-perceptual representations in this module (e.g., semantic) as it only has one layer. Moreover, while the differences between fast and slow learning rates are inspired by neurobiology and have been highly popular in the connectionist modeling literature, it is readily accepted that these models cannot be taken as direct neuroanatomical evidence and convergent research is needed. Indeed, the authors acknowledge that these models “are not intended as models of neuronal processing. Instead, they aim to explain cognitive processing in terms of statistical learning, using only a rough approximation of the actual learning mechanisms of the brain” (p. 370).

5.3 Frost and colleagues: Statistical learning in adults

Another highly interactionist model that proposes the modulation of perceptual systems during learning and experience is the multisensory model of statistical learning proposed by Frost, Armstrong, Siegelman, and Christiansen (2015). This model proposes that regions like the frontal cortex, the hippocampus and the striatum modulate relevant representations in perceptual regions during statistical learning. However, it is unclear whether this model is operational early in development as it is built exclusively from adult cognitive neuroscientistic findings (as virtually no developmental studies exist even in childhood: though see recent work from Finn, Kharitonova, Holtby, & Sheridan, 2019 linking the anatomical structure of the prefrontal cortex and the hippocampus to statistical learning abilities in childhood: 5–8.5 years).

6.1 A demonstration of fast, task/context-specific changes in infant perception

A major weakness in the current evidence for a top-down model of perceptual development is that we do not know if infants are able to rapidly modify their perception in a flexible and context-driven fashion. This type of rapid, flexible perceptual change is predicted to be available to young infants in a feedback/top-down model only and could provide a means by which feedback connections can drive broader changes in perceptual development. The current empirical findings are not sufficient: (1) They either show changes in perception as evidenced behaviorally but without a demonstration of flexibility or task-sensitivity (e.g., findings from Needham & Modi, 1999), or (2) They demonstrating rapid and specific changes but only neurally, no evidence from behavior (Emberson et al., 2015). While neural findings in particular are suggestive, they must be viewed with reservation because, given the neuroanatomical limitations of infants, it is possible that neural changes based on feedback are not sufficiently robust or rapid to modulate perception as evidenced in behavior.

6.2 Interaction with the development of long-range neural connectivity

Long-range neural connectivity exhibits a protracted developmental trajectory particularly in the white matter connections between disparate cortical regions with changes into adolescence (Barnea-Goraly et al., 2005). Thus, infants have limited neural connectivity which would presumably limit any potential long-range inter-regional interactions. However, the picture of connectivity in infancy has been changing with the increased focus on this topic in developmental cognitive neuroscience. Recent findings have demonstrated that while white matter does develop in a protracted development, there are rapid changes in white matter development in the first year or two of life (Dean et al., 2015). Moreover, while it was the dominant view from resting-state measures of functional connectivity, that infants had very poor long-range connectivity (Gao, Lin, Grewen, & Gilmore, 2017; Homae et al., 2010; Smyser, Snyder, & Neil, 2011), some recent studies have suggested that after a period of rapid development there is a core of adult-like connectivity by 6 months of age (Gilmore, Knickmeyer, & Gao, 2018; Sours et al., 2017). In general, it is important to determine how the use of feedback connections to modify perception relates to the development of long-range connectivity (determined both by measures of white matter and functional connectivity). It may be that there is a mutual relationship between these two developmental time-courses. Immaturity in long-range connectivity should constrain the use of feedback connections and the interactions between disparate regions of the brain but it may be that the interactions of disparate regions of the brain can help guide future refinement and development of long-range connectivity as suggested by Johnson (2011). Indeed, in adults, engagement of feedback connections has been shown to increase functional coupling across disparate regions of the brain (Al-Aidroos, Said, & Turk-Browne, 2012; Tompary, Duncan, & Davachi, 2015).

6.3 Determining the long-term impact of rapid changes in perception

While a feedback/top-down model of perceptual development predicts that infants would be capable of rapid changes in perception, perceptual development at a population-level emerges over months. If rapid top-down changes in perception can be demonstrated in infants, it will be necessary to reconcile perceptual changes at these two time-scales. In adults, Folstein et al. (2015) present evidence that rapid changes in visual perception, precipitated by top-down processes including learning and memory, persist for 11 days.

Similarly, it could be that very short periods of top-down modulation result in lasting changes in the cortex and need to be tested in infants. Relatedly, it may be possible that infants have many opportunities to engage in top-down modulation of their perceptual cortices based on their experience. These numerous, rapid, top-down changes in perception could build on each other and, ultimately, give rise to a broader developmental trajectory. Indeed, if changes in perception can occur arising from a few seconds to a few minutes of experience, infants are likely provided with many such bouts of experience that could support top-down engagement. An exciting area for future investigation is to determine the relationship between the structure of real-world infant experience (Fausey, Jayaraman, & Smith, 2016; Sugden & Moulson, 2018) and the learning abilities that infants demonstrate in the lab to determine how readily an infant could engage in top-down modulation of their perceptual systems in their daily lives.

6.4 Differentiate the sources of top-down modulation

A goal of the current conception of a feedback/top-down model is to bring attention to the need to carefully define and empirically evaluate the contribution of top-down modulation and neuroanatomical feedback connections in early perceptual development. However, to this end, the current model lumps many potentially disparate sources of top-down modulation into a single box within an over-simplified boxology (Fig. 1). However, if feedback/top-down models are ultimately favored over bottom-up ones, the differential development and contribution of these many potential sources of top-down modulation must be considered carefully and systematically. In the context of learning and memory, do different types of learning and memory systems differentially modulate visual perception in infancy either based on their developmental emergence, robustness, level of modulation in the visual system, etc.?

6.5 Why do developmental changes happen when they do?

As outlined earlier, exclusively feed-forward/bottom-up models predict a sequential type of development where development proceeds from lower-levels of vision to higher-levels, and then finally more cognitive influences can arise. In this model, it is very clear what develops but less clear how it develops. By contrast, in feedback/top-down models, many aspects of development can occur simultaneously as multiple regions of the brain and levels of processing interact. However, the assertion that every level develops together does not result in a mechanistic explanation of how perceptual development occurs. Indeed, specific testable hypotheses about what should develop when needs to determined and investigated. For example, if top-down modulation is available early in life and can be precipitated by short amount of experience, why do developmental changes in perception occur when they do at the population-level? For example, why does the other race effect (i.e., perceptual narrowing) occur from 6 to 10 months and not earlier? Are there constraints on the abilities of higher-level systems before that point? Is there an important change in input that allows the differential engagement of higher-level systems?