Visual Plasticity in Adulthood: Perspectives from Hebbian and Homeostatic Plasticity

Abstract

The visual system retains profound plastic potential in adulthood. In the current review, we summarize the evidence of preserved plasticity in the adult visual system during visual perceptual learning as well as both monocular and binocular visual deprivation. In each condition, we discuss how such evidence reflects two major cellular mechanisms of plasticity: Hebbian and homeostatic processes. We focus on how these two mechanisms work together to shape plasticity in the visual system. In addition, we discuss how these two mechanisms could be further revealed in future studies investigating cross-modal plasticity in the visual system.

Introduction

The nervous system changes in response to sensory input and environmental demands (Mayford and others 2012). Sensory input affects the nervous system most pronouncedly during early periods of development known as the “critical period.” The first systematic observation was based on the visual cortex of cats (Wiesel and Hubel 1963). When one eye was occluded for a few weeks after birth, the receptive field properties of neurons in the early visual cortex shifted in terms of which eye they preferentially responded to—a property known as “ocular dominance.” With the occlusion, ocular dominance columns that respond to stimulation from the open eye, expanded at the expense of the ocular dominance columns that responded to the closed eye. This effect was greatest between 4 and 8 weeks after birth and then reduced gradually until 3 months after birth. After this “critical period” was closed, monocular deprivation had little effect on the distribution of ocular dominance columns.

Similar effects are also found in the human visual cortex (Braddick and Atkinson 2011). When visual experience is suboptimal during early life due to eye problems such as congenital cataracts, amblyopia, and astigmatism, the visual cortex undergoes similar reorganization. For example, amblyopia, which arises as a result of abnormal visual input from one eye, may lead to imbalanced ocular dominance and poor visual acuity (Hensch and Quinlan 2018). Amblyopia can be treated by covering the preferred eye with a patch and forcing the brain to use signals from the amblyopic eye (Webber and Wood 2005). This treatment is most effective during early life and becomes less effective in adulthood, suggesting that the degree of plasticity is highest during early life but then decreases afterward (Fronius and others 2014).

However, studies over the past few decades revealed that the visual system preserves a substantial degree of plasticity even in adulthood (Castaldi and others 2020; Sagi 2011; Watanabe and Sasaki 2015). Its main evidence comes from visual perceptual learning, and visual deprivation ranging from short-term deprivation to complete blindness. A growing body of research suggests that experience and environmental demands such as visual training and visual deprivation change the visual system at multiple levels, from synapses (synaptic density, receptive field properties) to large-scale neural networks (cross-modal changes) (Karmarkar and Dan 2006).

Multiple mechanisms have been proposed to explain the continued plasticity of the adult visual system. Among various mechanisms, Hebbian and homeostatic plasticity appear to be two of the most influential cellular mechanisms. Hebbian plasticity refers to a synaptic change induced by the correlations between pre- and postsynaptic activity (Hebb and others 1994). For example, long-term potentiation (LTP) occurs when the presynaptic neuron fires before the postsynaptic neuron. If the postsynaptic neuron fires before the presynaptic neuron, long-term depression (LTD) occurs. By contrast, homeostatic plasticity is a mechanism that moves the neuronal system back toward its baseline after a perturbation—preventing the neuronal system from becoming hyper- or hypoactive, and stabilizing the neuronal activity (Turrigiano 2011; Turrigiano 1999). Both mechanisms have been proposed to play a critical role in regulating plasticity (Fox and Stryker 2017; Keck and others 2017).

Here, we focus on how the adult’s visual plasticity may be governed by Hebbian and homeostatic plasticity. Particularly, we concentrate on human neuroimaging studies. By nature, these human neuroimaging studies provide only indirect measures of changes in the synaptic activity as well as indirect evidence for either Hebbian or homeostatic plasticity due to their coarser spatiotemporal resolution (Poldrack 2000). However, despite the gap in resolution between systems and cellular level research, it is important to link research findings between these two different levels in order to better understand overall brain plasticity. Therefore, the current review aims to provide perspectives regarding how each plasticity mechanism may be utilized to regulate visual plasticity in the adult brain. To assess the plasticity mechanism in the visual system, we first introduce Hebbian and homeostatic plasticity in detail and summarize their supporting evidence from learning and memory research. This also includes a review of supporting evidence from non-visual learning due to the likelihood of shared underlying mechanisms. After reviewing two plasticity mechanisms and their supporting evidence from a wide range of learning types, we focus on the recent evidence of preserved plasticity in the adult visual system across visual perceptual learning, monocular and binocular deprivation in both short and long term. Finally, in each condition, we discuss how these findings reflect each of the two plasticity mechanisms.

Two Distinct Mechanisms for Plasticity

Hebbian Plasticity

Hebbian plasticity (Fig. 1A) proposes that the timing of pre- and postsynaptic activity determines the synaptic modification (Hebb and others 1994). For example, the synaptic efficacy that can be measured by changes in the synapse number or postsynaptic electrical currents, increases if the presynaptic cell firing is repeatedly followed by the postsynaptic cell firing. While it is challenging to relate this synaptic mechanism of Hebbian plasticity with cortical plasticity observed at the systems level, studies indicate that neuronal replay—a compelling candidate mechanism for brain plasticity that can be observed by neurophysiological or neuroimaging measures, is associated with Hebbian plasticity rules (O’Neill and others 2010). The details about how the replay is associated with Hebbian plasticity are discussed below.

Figure 1.

Hebbian plasticity versus homeostatic plasticity. (A) Hebbian plasticity suggests that the synaptic efficacy is determined by the timing of pre- and postsynaptic activity. For example, the long-term potentiation (LTP) is induced when the presynaptic activity is followed by the postsynaptic activity. On the other hand, the long-term depression (LTD) is triggered when the postsynaptic activity precedes the presynaptic activity. (B) Homeostatic plasticity proposes that perturbed neurons to either hyper- or hypoactive state are readjusted toward their baseline. One mechanism contributing to this process is synaptic downscaling, which can occur during sleep. When synapses become potentiated due to the encoding of information, the synaptic downscaling reduces the synaptic strength. Figure (A) was adapted from Andersen and others (2017). Figure (B) was reproduced from Turrigiano (2012).

Neuronal replay refers to the reexpression of the neural activity sequences exhibited during experience in the following rest or sleep. The first observation of replay comes from hippocampal place cells in rodents (Wilson and McNaughton 1994). During navigation, the hippocampal place cells are thought to form stronger connections with neighboring place cells that represent the next location where the animal heads to (Pfeiffer 2020). Such biased associations formed during navigation provides positive feedback to the system that can then promote later spontaneous replay (Chenkov and others 2017; Jackson and others 2006). Furthermore, replay itself is expected to reinforce these associations formed during navigation again through repeated co-firing of the involved cells (Sadowski and others 2016). As a result, it is posited that the synaptic connections of a replayed ensemble of neurons become strengthened, following Hebbian learning rules.

While neuronal replay has been detected primarily from neurophysiological recordings that have a fine spatial and temporal resolution, replay can be also detected at the level of blood oxygenation level–dependent (BOLD) signals measured by functional magnetic resonance imaging (fMRI) (Tambini and Davachi 2019; Wittkuhn and Schuck 2021). At the level of fMRI signals, replay is detected as a form of both sequential and nonsequential activity patterns, but the latter case is particularly referred to as reactivation (Tambini and Davachi 2019). Evidence of replay and reactivation was found in a wide range of brain areas, including the hippocampus (Schuck and Niv 2019), temporal lobe (Staresina and others 2013), higher-order association areas (Chelaru and others 2016; de Voogd and others 2016; Deuker and others 2013; Guidotti and others 2015; Schlichting and Preston 2014), and primary sensory areas (Bang and others 2018a). Critically, the strength of replay and reactivation was shown to predict subsequent performance improvements (Schlichting and Preston 2014; Tambini and Davachi 2013).

Replay and reactivation occur in temporal synchrony with sharp-wave ripple events (Buzsáki 2015; Lee and Wilson 2002) and sleep spindle activities (Bergmann and others 2012; Cairney and others 2018). The sharp-wave ripples are rapid bursts of synchronized neuronal activity generated by the hippocampus during both wakefulness and sleep (Buzsáki and others 1983) (Fig. 2B). The spindles are generated in the thalamic reticular nucleus and relayed to the cortex during sleep (De Gennaro and Ferrara 2003) (Fig. 2B). In the brief time window of hippocampal sharp-wave ripples during wakefulness, brain activity patterns representing previous experiences are reinstated in the cortical areas, suggesting that the hippocampal sharp-wave ripples trigger memory retrieval and orchestrate the reinstatement of cortical representations (Norman and others 2019). Similarly, during sleep, replay and reactivation co-occur with the hippocampal sharp-wave ripples (Buzsáki 2015; Lee and Wilson 2002) and spindles (Bergmann and others 2012; Cairney and others 2018). Particularly, the spindle-coupled reactivation appears in topographically restricted cortical areas involved in learning (Bergmann and others 2012; Cairney and others 2018).

Figure 2.

(A) Sleep has two types, non–rapid eye movement (NREM) sleep, and rapid eye movement (REM) sleep. The sleep cycle of alternating NREM and REM sleep is 90 to 120 minutes and appears four to five times in overnight sleep. During the first half of sleep, NREM sleep prevails, whereas in the second half of sleep, REM sleep occupies more time. NREM sleep includes light sleep stages 1 and 2, and slow-wave sleep (SWS) consisting of deeper stages 3 and 4. (B) The most prominent brain oscillations during NREM sleep include sharp wave ripples, spindles, and slow waves. Sharp-wave ripples are high-frequency waves (100–300 Hz) generated in the hippocampus during stages 3 and 4. Spindles are waxing and waning waves of 10 to 16 Hz, generated in the thalamus during stages 2 to 4. Spindles are relayed to the cortex through a thalamocortical loop. Slow waves are high amplitude oscillations with 0.5 to 4 Hz, appearing in the cortex during stages 3 and 4. Slow waves consist of a down-state with synchronized neuronal silence and an up-state with widespread neuronal firing. Theta waves are oscillations with 4 to 7 Hz that appear in the hippocampus and cortex during REM sleep. Figure is adapted from Diekelmann and Born (2010).

In hippocampus-dependent learning, it is widely agreed that new memories are temporally stored in the hippocampus but then gradually transferred into the neocortex for long-term storage (Diekelmann and Born 2010). Specifically, a temporal coupling of sharp-wave ripples, spindles, and slow waves, which are synchronized oscillatory activity at 0.5 to 4 Hz has been proposed as a main mechanism supporting this shift between hippocampus and neocortex during sleep (Diekelmann and Born 2010; Klinzing and others 2019). The sharp-wave ripples accompanying memory replay/reactivation in the hippocampus appear to be nested within the excitable troughs of the spindles (Siapas and Wilson 1998). The spindles then appear to be nested in the excitable up-state of the cortical slow waves (Clemens and others 2007).

Further evidence supporting the notion that replay and reactivation are tightly associated with Hebbian plasticity comes from neurophysiological studies. For example, reactivated cell firing patterns were found to induce LTP in the hippocampus, and such induction of LTP required sharp-wave ripple events (Sadowski and others 2016). Similarly, spindle-like stimulation was shown to induce LTP in the cortical pyramidal cells (Rosanova and Ulrich 2005). This line of studies suggest that replay/reactivation, sharp-wave ripples, and spindles may be indicative of Hebbian plasticity.

Homeostatic Plasticity

Homeostatic plasticity (Fig. 1B) refers to the global readjustment of all synapses toward baseline (Turrigiano 2011; Turrigiano 1999). Such global readjustment allows for the neural system to maintain an overall balance of network excitability that is required for the optimal performance of neurons. For example, if the synapses are potentiated due to information encoding, all synapses undergo downscaling processes. On the other hand, if the synaptic connections are weakened from sensory perturbation, all synapses become upscaled. Its evidence was first observed from cell culture studies (Turrigiano 2011). Pharmacological manipulations that increase neuronal activity led to synaptic downscaling, whereas drugs that inhibit neuronal activity resulted in upscaling (Turrigiano and others 1998). Similar observations were made from in vivo studies where sensory deprivation led to synaptic upscaling (Goel and Lee 2007; Wallace and Bear 2004).

This concept of synaptic scaling was then incorporated into the synaptic homeostasis hypothesis (Frank 2012; Tononi and Cirelli 2006). According to the synaptic homeostasis hypothesis, the synaptic connections are potentiated during wakefulness and then globally downscaled via slow wave activity during non–rapid eye movement (NREM) sleep (Fig. 2).

A number of studies provide results that are consistent with this synaptic homeostasis model. For example, the synaptic efficacy in the brains of rodents (Liu and others 2010; Vyazovskiy and others 2008) and the size and number of synapses in Drosophila neurons (Bushey and others 2011) increase after wakefulness but then decrease after sleep. The involvement of the slow wave activity in synaptic downscaling is supported by the finding that depolarizing currents at around 1 Hz lead to LTD (Czarnecki and others 2007) and that presynaptic stimulation during Up-states of slow waves leads to LTD unless it is followed by postsynaptic firing within a short time window (Gonzalez-Rueda and others 2018). Additionally, the slow wave activity is sensitive to the usage of neurons during wakefulness (Huber and others 2006; Kattler and others 1994; Vyazovskiy and others 2000). For example, immobilizing an arm during wakefulness decreases the slow wave activity within the corresponding motor cortex during NREM sleep (Huber and others 2006). In contrast, training individuals on a motor task increases slow wave activity in motor areas during NREM sleep (Tamaki and others 2013).

While the synaptic homeostasis hypothesis proposes the involvement of slow wave activity only in the process of downscaling, accumulating studies show that theta activity at 4 to 7 Hz, a prominent feature of REM sleep, is involved in synaptic downscaling as well (Born and Feld 2012). For example, in the cycle of NREM-REM-NREM sleep, the firing rates of neurons significantly decreased during REM sleep (Grosmark and others 2012). Critically, a reduction of firing rate during REM sleep was correlated with the power of theta activity. Consistent with this finding, other studies showed that a disruption of theta activity results in impaired memory (Boyce and others 2016; Swift and others 2018) and oversized place fields of hippocampal neurons (Swift and others 2018). Oversized place fields suggest that theta disruption prevented synaptic downscaling, which would otherwise occur via theta activity while optimizing the size of place fields of hippocampal neurons (Pereira and Lewis 2020).

Hebbian and homeostatic plasticity frequently operate in opposite directions (Fox and Stryker 2017; Keck and others 2017). For example, Hebbian plasticity suggests that when neuronal activity increases, this increase in synaptic gain provides positive feedback to the system, which again enhances the probability that the synaptic gain is further increased. On the other hand, homeostatic plasticity provides negative feedback such that increased neuronal activity is downscaled in the following periods. Despite such seemingly opposite directions, it is generally accepted that both mechanisms act in concert to maintain a balance between plasticity and stability (Fox and Stryker 2017; Keck and others 2017). For instance, homeostatic plasticity prevents the saturation of synapses that could be otherwise induced by Hebbian plasticity. Critically, the relative differences of synaptic strength generated by Hebbian plasticity are thought to be preserved. One possible mechanism enabling this is the protection against LTD by postsynaptic spikes (Gonzalez-Rueda and others 2018). When the presynaptic activation during up-states of slow waves is followed by postsynaptic firing, the synaptic connection is protected from LTD. This mechanism allows selective downscaling by which synaptic connections that were strengthened during wakefulness are protected from weakening. It suggests that only weak connections are pruned whereas strong connections survive from the synaptic downscaling. Such different effects of downscaling on synaptic connections are thought to enhance the signal-to-noise ratio of the encoded information (Gonzalez-Rueda and others 2018). This topic regarding how Hebbian and homeostatic plasticity are integrated in the brain is discussed intensively in other articles (Fox and Stryker 2017; Keck and others 2017).

Plasticity in Visual Perceptual Learning

Visual perceptual learning refers to performance improvements following visual training (Sagi 2011; Watanabe and Sasaki 2015). A number of studies demonstrate that visual learning resulting from training is highly specific to the trained visual features, such as orientation (Crist and others 1997), spatial frequency (Fiorentini and Berardi 1980), motion (Ball and Sekuler 1987), texture (Karni and Sagi 1991), and position (Fahle 2004). Such specificity is often interpreted as evidence of plasticity in the early stages of visual processing where receptive fields have fine selectivity for visual features (Sagi 2011; Watanabe and Sasaki 2015). However, the brain areas where changes occur as a result of visual training have been highly debated. In this section, we briefly introduce four major changes (representational changes, recurrent changes, read-out changes, feedback changes) that can occur with visual training (Seitz 2020). Then we discuss the evidence of changes in the visual cortex in light of the aforementioned four categories. Furthermore, we discuss how the brain changes observed in the visual cortex may support Hebbian or homeostatic plasticity.

Plasticity after visual training can be largely grouped into four domains (Seitz 2020). Representational changes describe that receptive field properties of neurons in the early visual cortex can change as a result of visual training. For example, the orientation-selective neurons in the primary visual cortex (V1) can sharpen their tuning curve around their preferred orientation (Schoups and others 2001) or shift their preference to a different orientation (Dragoi and others 2000) with visual training. Recurrent changes refer to alterations of the connections in the superficial layers of the cortex with training. Critically, it was proposed that recurrent processing is supported by long-range horizontal connections and that recurrent changes could contribute to representational changes in the early visual cortex (Jia and others 2020). For example, the tuning curve of orientation-selective neurons are known to be shaped by iso-orientation surround suppression, that is the inhibition of neurons that prefer the same orientation across columns (Shushruth and others 2012). Thus, if visual training alters the horizontal inhibitory connections in the early visual cortex, the tuning curves of the orientation-selective neurons are expected to change as well. Read-out changes describe that the connection weights in feedforward connectivity from the early visual cortex to decision areas are changed as a result of training (Dosher and Lu 2017). This model does not support representational changes in V1, but rather proposes that the reweighting between early visual cortex and decision areas improves the read-out process of the sensory signals. Last, feedback changes describe that the feedback connections from decision areas to the visual cortex are changed as a result of training (Byers and Serences 2012).

Although the question of where in the brain as well as what specific changes visual training induces has been heavily debated, prior electrophysiological studies have presented mixed results. For example, some studies reported training-related changes in V1 (Hua and others 2010; Schoups and others 2001; Yan and others 2018), V4 (Adab and Vogels 2011; Yang and Maunsell 2004), and middle temporal area (MT) (Zohary and others 1994), whereas other studies found little to no changes in early (e.g., V1) and mid-level (e.g., V4 and MT) visual areas (Ghose and others 2002; Law and Gold 2008). Furthermore, some behavioral studies observed that learning is transferable under specific conditions (Harris and others 2012; Tan and others 2019; Wang and others 2016; Xiao and others 2008), whereas others observed that learning is specific for the trained stimulus (Ball and Sekuler 1987; Crist and others 1997; Fahle 2004; Fiorentini and Berardi 1980) and even the trained eye (Karni and Sagi 1991). Such divergent results are likely to reflect that plasticity resulting from visual training involves multistage mechanisms that may vary depending on the tasks and stimuli used during training as well as the temporal phases of the training (Shibata and others 2014). Readers who are interested in how these divergent results may be reconciled in a unified framework may refer to the following theory article (Shibata and others 2014).

While it is critical to reconcile divergent results in visual perceptual learning, here we first focus on the evidence of changes in the visual cortex of humans. Then we discuss these findings in light of the aforementioned four domains.

In human neuroimaging studies, BOLD activity was shown to increase in response to the presentation of trained stimuli in V1 (Furmanski and others 2004; Schwartz and others 2002; Walker and others 2005; Yotsumoto and others 2008) and even in the lateral geniculate nucleus (LGN), which relays visual input from the retina to V1 (Yu and others 2016). This brain activity increase was observed to disappear in a later phase of extensive learning where the training was conducted for more than 14 days (Jehee and others 2012; Yotsumoto and others 2008). However, despite such return of BOLD amplitude activity to baseline in a later phase of learning, the neural representation of the trained orientation, as measured by orientation discriminability (d′), was still shown to be enhanced in V1 as well as V2-V4 (Jehee and others 2012). This improvement of neural discriminability of the trained orientation in V1 also predicted the behavioral enhancement in the task.

Building on this, a recent study investigated the laminar-specific changes in the visual cortex and the decision area—the intraparietal sulcus (IPS), using ultra-high-field fMRI (Jia and others 2020). Specifically, this study observed learning-specific neural changes indicated by the enhanced neural representation of the trained orientation within the superficial layers of V1 and middle layers of IPS. The superficial layers of V1 are characterized by horizontal connections supporting recurrent processing as well as ascending feedforward projections to higher areas. Whereas the middle layers of IPS receive ascending feedforward signals from lower areas. In addition to changes within V1 and IPS, neural changes were observed within feedforward connections from the superficial layers of V1 to the middle layers of IPS. These findings suggest that visual training changes the recurrent processing within the superficial layers of V1 and strengthens the feedforward connectivity from V1 to IPS. Therefore, these results can be interpreted with the combination of representational changes, recurrent changes, and read-out changes. However, they do not rule out the possibility that visual perceptual learning involves changes in the feedback signals. This is because fMRI signals are inherently influenced by both feedforward and feedback signals. Indeed, some studies using other types of visual tasks demonstrated changes within decision areas such as the IPS (Kahnt and others 2011; Law and Gold 2008; Shadlen and Newsome 2001) and the dorso-lateral prefrontal cortex (Heekeren and others 2004) as a result of training.

While it is difficult to disentangle changes between representation, recurrent connections, read-out processes, and feedback processes from the observed changes in neuroimaging measures, it is likely that they reflect mixtures of these four domains. Thus, rather than attempting to disentangle four possible changes apart, here we focus on the temporal dynamics of the neural changes observed in the visual cortex. Particularly, we review the neural mechanisms occurring after the offset of training and discuss these changes in the perspective of memory consolidation. The case of fast learning occurring during a task session is thus not covered in this article but elsewhere (Fahle and others 1995; Poggio and others 1992). Finally, we discuss how these neural changes may reflect Hebbian or homeostatic plasticity.

Consolidation of Visual Perceptual Learning

The neural circuits involved in visual perceptual learning undergo a consolidation process after the offset of training (Sagi 2011). The consolidation process refers to the mechanism that transforms a newly learned memory into a long-lasting stable memory.

There are at least two different benefits from the consolidation process in visual perceptual learning (Sagi 2011). The first benefit is resilience to interference. Visual perceptual learning is fragile against interference shortly after the offset of training, but then becomes resilient to interference over the course of time. If new learning follows shortly after the offset of the first learning, the initial learning is easily disrupted, resulting in no performance improvement (Hung and Seitz 2011; Seitz and others 2005; Shibata and others 2017; Yotsumoto and others 2009). The duration of time needed for the stabilization was found to be 1 hour in a hyperacuity task (Hung and Seitz 2011; Seitz and others 2005), whereas it was 3.5 hours for an orientation detection task (Shibata and others 2017). Such stabilization is shown to occur not only during wakefulness but also during sleep (Tamaki and others 2020).

The second benefit of consolidation is offline performance improvement. Individuals show enhanced performance on the trained visual task after a period of wakeful rest (Censor and others 2006; Karni and Sagi 1993) or sleep (Stickgold and others 2000a; Stickgold and others 2000b), even though there is no further practice in between the sessions. Interestingly, these two benefits of consolidation seem to occur spontaneously. However, the consolidation process itself requires following awake and/or sleep periods to unfold, likely involving Hebbian and homeostatic plasticity. This is evidenced by the fact that learning is abolished when the neural processes are disrupted by transcranial magnetic stimulation (TMS) after the offset of training (Bang and others 2019).

Consolidation during Wakefulness

Which neuronal mechanisms underlie the consolidation process during wakefulness? Here, we discuss the involvements of the excitatory and inhibitory (E/I) balancing dynamics and awake reactivation/suppression in the consolidation of visual perceptual learning.

Consolidation during Wakefulness: Neurochemical Changes (E/I Ratio).

The balance between excitation and inhibition in the early visual cortex (V1, V2, and V3) is involved in the consolidation process of visual perceptual learning (Bang and others 2018b; Shibata and others 2017). This balance between excitation and inhibition is evaluated as an E/I ratio—comparing the concentrations of excitatory (i.e., glutamate or glutamate/glutamine complex [Glx]) to inhibitory (i.e., gamma-aminobutyric acid [GABA]) neurochemicals from a single voxel of approximately 2 × 2 × 2 cm in size positioned in the early visual cortex using proton magnetic resonance spectroscopy (MRS). MRS is capable of detecting the neurochemicals in the local region of the brain non-invasively, thus providing a powerful tool to track neurochemical changes. Using this technique, studies showed that when visual learning was in a vulnerable state to interference (e.g., immediately after the offset of training), the E/I ratio in the early visual cortex was significantly higher than baseline, putting the visual cortex in an excitatory-dominant state (Shibata and others 2017). However, this increased vulnerability and the E/I ratio tapered off following 3.5 hours of consolidation after training. These results suggest that the plastic state of visual learning is associated with a higher E/I ratio in the early visual cortex and that the consolidation process, which stabilizes this learning unfolds during which the elevated E/I ratio goes back to baseline.

While the degree of plasticity increases immediately after the offset of training and gradually decreases toward baseline during the following hours, overlearning appears to be an exception to this trend (Fig. 3). Overlearning, which is defined as continued practice after the performance enhancement has plateaued, triggers abrupt decreases in the E/I ratio before a gradual return toward baseline (Shibata and others 2017). In addition, overlearning hyper-stabilizes visual perceptual learning, that is visual learning becomes not only resilient against interference but even disrupts its following new learning. Such hyperstabilization and abrupt decreases in the E/I ratio in the early visual cortex tapered off within 3.5 hours posttraining. Here, it remains to be elucidated whether the return of decreased E/I ratio back to baseline in the case of overlearning is associated with the consolidation process.

Figure 3.

(A) Schematic illustration showing how two different Gabor orientations illicit different multivoxel patterns in the early visual cortex (V1, V2, V3). Each pixel in the activity represents each voxel from functional magnetic resonance imaging (fMRI) data. Each pixel’s color represents the blood oxygenation level–dependent (BOLD) activity value. (B and C) Schematic illustration of awake reactivation in the primary visual cortex (V1). (B) Shortly after visual training (≈40 minutes), the brain activity patterns in V1 appear similar to the recently trained orientation. (C) Brief exposure to a task with novel (untrained) orientation (≈5 minutes) leads to awake reactivation in V1, as in training. (D) Schematic illustration of awake suppression in V1, V2, and V3. Brief exposure to a task with familiar (previously trained) orientation (≈5 minutes) triggers awake suppression. The neural activity patterns associated with that stimulus are suppressed across V1, V2, and V3.

Converging studies propose that the consolidation process can occur more than once. For example, if subjects are provided with a brief task that they were previously trained on, visual learning becomes vulnerable again, and subsequently enters the reconsolidation process. (Amar-Halpert and others 2017; Bang and others 2018b; Shmuel and others 2021). This re-consolidation process then can result in offline performance improvement (Amar-Halpert and others 2017; Censor and Sagi 2008). The E/I ratio dynamics during reconsolidation showed similar patterns as with consolidation (Bang and others 2018b). Here the E/I ratio exhibits an immediate increase followed by a gradual decrease toward baseline levels after retrieval.

Overall, this line of MRS studies suggest that the consolidation process of visual learning involves E/I ratio changes in the early visual cortex. These studies did not observe significant changes in glutamate or GABA in isolation, which may be due to their related nature as most GABA is synthesized from glutamate (Petroff 2002).

The link between the high degree of plasticity and elevated E/I ratios observed in the above MRS studies is consistent with animal studies showing that previously closed ocular dominance plasticity in adult rodents could be recovered by reducing GABA-mediated inhibition in the visual cortex (Harauzov and others 2010; Maya Vetencourt and others 2008). Thus, GABA has been suggested as a potential brake for plasticity (Hensch and Quinlan 2018). GABA exists in two intracellular pools— a cytoplasmic pool and a vesicular pool (Martin and Rimvall 1993). While cytoplasmic GABA is considered to be involved in metabolism, vesicular GABA observed within the presynaptic boutons plays a role in inhibitory synaptic neurotransmission (Martin and Rimvall 1993). Additional GABA observed in the extracellular pool is known to play a neuromodulatory role in cortical inhibition (Belelli and others 2009). While conventional MRS cannot distinguish between these different pools, it would be interesting to see which GABA pools are affected by visual training in future studies using advanced techniques such as diffusion-weighted MRS (Ronen and Valette 2014).

Consolidation during Wakefulness: Change in BOLD Activity Patterns.

Visual training leads to a reactivation of the trained stimulus in V1 (Fig. 3A and B) (Bang and others 2018a). When subjects were trained on a visual feature such as a Gabor orientation, the multivoxel pattern of BOLD activities representing the trained orientation was reexpressed in V1 shortly after the training offset. Critically, subjects who showed a stronger reactivation (as measured by an increase in decodability of the trained stimulus in V1 after training) displayed greater performance improvement later.

It should be considered that reactivating all visual information is costly since the visual cortex is continuously bombarded with visual information relayed from the retina during wakefulness. Thus, the brain has likely developed a mechanism that selects which experience should be reactivated and which experience should be ignored. Indeed, recent studies showed that the brain can distinguish between such stimuli when engaging in reactivation. For example, the brain has been shown to prioritize new information that has been remembered less well (Schapiro and others 2018) or information paired with a reward (Gruber and others 2016; Murty and others 2017) or shock (de Voogd and others 2016) for reactivation.

Based on these findings, a recent study examined whether awake reactivation occurs similarly for new vs extensively trained visual features using Gabor orientations (Bang and Rahnev 2021). Here, exposure to a novel orientation led to awake reactivation (Fig. 3C) while exposure to an extensively trained orientation led to the opposite effect, with brain activity patterns associated with the familiar orientation being suppressed (Fig. 3D). Interestingly, awake reactivation was the strongest in V1, whereas awake suppression had a similar strength across V1, V2, and V3. This pattern is consistent with the notion that awake reactivation may be a local process in V1, whereas awake suppression may be a top-down process. Future studies are needed to elucidate such directions.

In the above study, both reactivation and suppression were measured while subjects performed a fixation task on a gray background where a target did not spatially overlap with the location of the trained stimulus. Thus, one interesting question that arises from this study is how strong reactivation and suppression may be in normal daily life where visual information is continuously relayed from the retina.

The effect of awake suppression on behavior remains an open question as well. We expect that awake suppression may have different effects from awake reactivation, which predicts performance improvement. This is because awake suppression was observed only for extensively trained orientations and a prior study showed that overtraining makes the visual learning hyperstabilized (Shibata and others 2017). Thus, the effect of awake suppression on behavior may be related to hyperstabilized state of visual learning, however, this hypothesis warrants further examinations in future studies.

Taken together, the aforementioned studies suggest that the E/I dynamics (Bang and others 2018b; Shibata and others 2017) and awake reactivation (Bang and Rahnev 2021; Bang and others 2018a) are associated with the consolidation process. Here, it should be noted that a high E/I ratio and reactivation were observed in a similar time window shortly after the offset of training. Elevated E/I ratios suggest that the early visual cortex including V1, V2, and V3 are in an excitatory dominant state. The reactivation of trained orientations in V1 is likely to increase the excitation and further strengthen the synaptic connections of involved neurons (Pfeiffer 2020; Sadowski and others 2016). Therefore, it is possible that both elevated E/I ratios and reactivations are linked together in a loop such that reactivation increases E/I ratios and increased excitation promotes reactivation. We speculate that these two may reflect Hebbian plasticity because the co-firing of involved neurons that are suggested to underlie the reactivation are thought to increase the synaptic efficacy (Pfeiffer 2020; Sadowski and others 2016). On the other hand, to date, we lack the evidence that the dynamics of the E/I ratio and reactivation may be associated with the synaptic downscaling, a key process in the homeostatic plasticity.

Consolidation during Sleep

Which neuronal mechanisms underlie the consolidation process during sleep? Here, we focus on the dynamics of the E/I ratio and spontaneous brain oscillations during sleep.

Consolidation during Sleep: Neurochemical Changes (E/I Ratio).

Sleep is categorized by NREM and REM sleep (see Fig. 2 for the sleep structure and the spontaneous brain oscillations during sleep). NREM sleep initiation is gradual, consisting of sleep stages 1, 2, and slow-wave sleep (SWS). NREM sleep appears dominant during the first half of sleep, whereas REM sleep prevails during the second half of sleep. Both NREM and REM sleep have been proposed to consolidate visual perceptual learning. For example, the offline performance improvement requires sleep within 30 hours of training (Stickgold and others 2000a) and is correlated with the duration of SWS (the deepest phase of NREM sleep) in the first quarter of sleep, as well as the duration of REM sleep in the last quarter of sleep (Stickgold and others 2000b). A brief (60–90 minutes) nap appears to benefit the visual perceptual learning as well, by improving the offline performance improvement (Mednick and others 2003) and preventing performance deterioration that otherwise emerges with an excessive amount of training in a single day (Mednick and others 2002).

Building on these studies, it was questioned whether NREM and REM sleep have different roles in consolidation. To address this question, previous studies attempted to manipulate sleep by disrupting either NREM or REM sleep (Karni and others 1994) or by restricting sleep by the first NREM sleep (Bang and others 2014; Mednick and others 2002). These studies observed mixed results with some findings supporting that SWS, the deepest phase of NREM sleep is sufficient for the offline performance (Bang and others 2014; Mednick and others 2002) and others supporting that REM sleep is required for learning (Karni and others 1994).

While the role of NREM and REM sleep remains controversial in terms of offline performance improvement, a recent study further suggested that NREM and REM sleep have two complementary mechanisms in the consolidation of visual learning (Tamaki and others 2020). Specifically, this study indicates that NREM sleep contributes to improving performance whereas REM sleep plays a role in stabilizing visual learning. Its supporting evidence comes from behavioral measures and E/I ratio changes of the early visual cortex measured by MRS. During NREM sleep, the E/I ratio increases, driven by a reduction in GABA. This increased E/I ratio during NREM sleep predicts subsequent offline performance improvement. In contrast, during REM sleep, the E/I ratio decreases below baseline, driven by a reduction of Glx (Fig. 4). This decreased E/I ratio during REM sleep predicts the degree to which presleep learning is resilient against interference from postsleep visual learning.

Figure 4.

Effects of non–rapid eye movement (NREM) and rapid eye movement (REM) sleep on dendritic spines and the excitatory/inhibitory (E/I) ratio. During NREM sleep after task learning, new spines are formed, and the E/I ratio increases within the involved brain areas. Such increases in the E/I ratio are driven by reduced inhibition. In the following REM sleep, some of the newly formed spines are pruned and the E/I ratio decreases. Such decreases in the E/I ratio are driven by reduced excitation. Figure is adapted from Sun and others (2020).

The above results are consistent with a recent animal study that investigated spine morphology at the cellular level during sleep. This animal study observed a surge of new spines formed during NREM and the subsequent pruning and strengthening of selective spines during REM sleep in the dendrites of pyramidal neurons after learning (Fig. 4) (Li and others 2017). Based on these results, it was proposed that increased E/I ratio driven by reduced GABA during NREM sleep (Tamaki and others 2020) may be related with formation of new spines, whereas decreased E/I ratio due to reduced Glx during REM sleep (Tamaki and others 2020) may be associated with selective pruning of spines in the dendrites of neurons (Pereira and Lewis 2020). However, such an association is still speculative and needs to be further addressed in future studies.

Consolidation during Sleep: Neurophysiological Change (Spontaneous Oscillations).

It is widely accepted that sharp-wave ripple activity (Buzsáki 2015; Lee and Wilson 2002), spindle activity (Bergmann and others 2012; Cairney and others 2018), and slow-wave activity during NREM sleep (Tamaki and others 2013) as well as theta activity during REM sleep (Boyce and others 2016; Nishida and others 2009) are involved in memory consolidation. Among these oscillations, spindle and theta activities were shown to increase within the trained region of early visual areas (V1, V2, V3) during NREM and REM sleep, respectively (Bang and others 2014; Tamaki and Sasaki 2020). The strength of spindle activity predicted offline performance improvement (Bang and others 2014; Tamaki and Sasaki 2020) and the power of theta activity predicted the degree that presleep learning is protected against interference from postsleep learning (Tamaki and others 2020). Interestingly, the power of theta activity was negatively correlated with the E/I ratio during REM sleep, suggesting that the higher theta power is associated with lower E/I ratios in the early visual cortex (Tamaki and others 2020).

It is noteworthy that during NREM sleep, the power of the spindle activity and E/I ratio increased within the early visual cortex (Bang and others 2014; Tamaki and Sasaki 2020; Tamaki and others 2020). This suggests a link between the spindle event and elevated E/I ratio during NREM sleep. One possibility is that reactivation of the trained stimulus is coupled with spindle events and that such spindle activity accompanies the formation of new spines resulting in a higher E/I ratio. Indeed, reactivation occurs in temporal synchrony with spindle events (Bergmann and others 2012; Cairney and others 2018). Furthermore, the spindle activity accompanies the highest activity of the pyramidal cells, which may reflect the upregulation of synapse formation (Niethard and others 2017). Enhanced activity of pyramidal cells during spindle activity is then in line with the elevated E/I ratio that was observed in the early visual cortex during NREM sleep (Tamaki and others 2020). This explanation favors Hebbian plasticity, which proposes that synaptic connections formed during wakefulness provide positive feedback to the system, resulting in further strengthening via LTP. Indeed, spindle-like stimulation and reactivated cell firing were shown to induce LTP (Rosanova and Ulrich 2005; Sadowski and others 2016).

The increase of theta activity (Tamaki and Sasaki 2020) and decrease of E/I ratio (Tamaki and others 2020) in the early visual cortex during REM sleep suggest a connection between the two during REM sleep. A compelling explanation is that synaptic downscaling is carried via theta activity in the early visual cortex during REM sleep and that this synaptic downscaling results in synaptic pruning and a decreased E/I ratio (Pereira and Lewis 2020; Tamaki and others 2020). This explanation is supported by studies showing that theta oscillation is involved in synaptic downscaling (Grosmark and others 2012; Swift and others 2018) and that increased theta activity predicts lower E/I ratios during REM sleep (Tamaki and others 2020). The decreased E/I ratio during REM sleep has been proposed to reflect pruning of glutamatergic excitatory synapses during REM sleep (Tamaki and others 2020). Thus, a hypothetical link between theta activity, synaptic downscaling, synaptic pruning, and a decreased E/I ratio during REM sleep favors homeostatic plasticity, which proposes that downscaling results in a pruning of weak connections.

While spindle and theta activities appear to be involved in the consolidation of visual perceptual learning, the role of slow-wave activity remains elusive. One study found that the number of slow waves initiated in the occipital areas was correlated with performance improvements (Mascetti and others 2013). However, other studies did not find any changes in the slow wave activity after visual training (Bang and others 2014; Tamaki and Sasaki 2020).

One may wonder if the hippocampal sharp-wave ripples are involved in visual perceptual learning. While the sharp-wave ripples are commonly observed for hippocampus-dependent memory, there is evidence that visual perceptual learning occurs independently of the hippocampus (Fahle 2004). Thus, it appears less likely that visual perceptual learning involves the sharp-wave ripples.

In summary, studies in visual perceptual learning demonstrate that the neural circuits involved in training go through a consolidation process during the following wakefulness and sleep periods after the offset of training. The consolidation process involves E/I dynamics, reactivation of trained features, as well as spindle and theta activities. Specifically, the increased E/I ratio and reactivation of trained features observed within the same wakefulness period favor Hebbian plasticity. In addition, the increased E/I ratio and enhanced spindle activity observed during the following NREM sleep period are in line with Hebbian plasticity. By contrast, the decreased E/I ratio and increased theta activity observed during REM sleep favor homeostatic plasticity. This suggests that Hebbian plasticity is dominant during wakefulness and NREM sleep, while homeostatic plasticity prevails during REM sleep in the case of visual perceptual learning. However, the full extent of these processes remains unknown and open for many possibilities. For example, homeostatic plasticity may occur beyond REM sleep or both Hebbian and homeostatic plasticity may coexist during wakefulness and sleep. Overall, the evidence appears to support the notion that both Hebbian and homeostatic plasticity are involved in shaping the brain plasticity resulting from visual perceptual learning.

Visual Deprivation in the Sighted

Visual deprivation during adulthood induces neuronal changes in the visual cortex. Here we review the effects of visual deprivation in two categories: monocular and binocular visual deprivation. In each category, we discuss its short- and long-term effects. Particularly, we provide an overview of the long-term effects of monocular and binocular deprivation with the cases of amblyopia and total blindness, respectively. Furthermore, we discuss the roles of Hebbian and homeostatic plasticity under such conditions.

Monocular Visual Deprivation

Short-term monocular deprivation shifts eye dominance toward the deprived eye (Fig. 5A). This is shown by binocular rivalry, in which visual perception “flips” between different images shown to each eye. Normally, the time on each image is comparable, implying equal eye dominance. However, studies have shown that after 2.5 hours of monocular deprivation, perception becomes biased toward the deprived eye for twice as long as the non-deprived eye (Lunghi and others 2011). This effect remains for 1.5 hours for monochromatic vision (Lunghi and others 2011), or 3 hours for chromatic vision (Lunghi and others 2013).

Figure 5.

(A) Short-term monocular deprivation biases perception toward the deprived eye. When two different images are presented to each eye, the perception alternates between two images with comparable time. However, after monocular deprivation, the perception is dominated by the image presented to the deprived eye. (B) Short-term monocular deprivation alters blood oxygenation level–dependent (BOLD) activity in the primary visual cortex (V1). The percentage change of BOLD activity induced by stimulation of the deprived eye is enhanced but that of the non-deprived eye is reduced in V1 after monocular deprivation. (C) The eye dominance change is correlated with the deprivation index assessed with BOLD activity in V1. (D) The eye dominance change is correlated with the change in gamma-aminobutyric acid (GABA) amount in V1. Figure B and C is adapted from Binda and others (2018). Figure D is adapted from Lunghi and others (2015b).

Eye dominance changes induced by monocular deprivation are likely driven by altered brain activities in the visual cortex. The amplitude of the C1 component of visual evoked potentials, alpha band peaks, and V1 BOLD activities increase for deprived eye stimulation but decrease for the non-deprived eye stimulation (Fig. 5B) (Binda and others 2018; Lunghi and others 2015a). In particular, the deprivation effect on BOLD responses in V1 is correlated with eye dominance change as measured by binocular rivalry (Fig. 5C). Relatedly, V1 voxels preferring non-deprived eye stimulation shift their preference toward the deprived eye, suggesting that ocular dominance may be altered. These results are in line with homeostatic plasticity and can be interpreted such that the brain enhances the signals from the deprived eye to maintain the balance of the network excitability. On the other hand, an alternative interpretation was also raised, which is that increases of the signals from the deprived eye are due to a release from adaptation (inhibition) (Binda and others 2018; Zhang and others 2009). While it remains unclear whether monocular deprivation effect is due to homeostatic plasticity or a release from adaptation, both mechanisms seem to yield similar output, that is, keeping the overall brain activity constant (Binda and others 2018).

Eye dominance changes due to monocular deprivation involve neurochemical changes in the visual cortex. Monocular deprivation reduces the amount of GABA in V1 and this reduction is correlated with altered eye dominance favoring the deprived eye (Fig. 5D) (Lunghi and others 2015b). Furthermore, norepinephrine, one of the excitatory neurotransmitters, is thought to be involved in monocular deprivation. This is because the amplitude of slow pupil oscillations (hippus) increases in proportion to the eye dominance changes in the case of monocular deprivation (Binda and Lunghi 2017). It should also be noted also that norepinephrine plays a role in controlling pupil dilations (Joshi and others 2016) and ocular dominance plasticity (Kasamatsu and others 1979).

Multisensory interactions are affected by monocular deprivation as well (Lo Verde and others 2017; Opoku-Baah and Wallace 2020). For example, when haptic stimulation is provided during binocular rivalry, haptic and visual processes interact with each other such that the probability of seeing a congruent visuo-haptic stimulus increases. However, after 2.5 hours of monocular deprivation, this observed effect of touch disappears for the deprived eye, whereas the effect remains the same for the non-deprived eye (Lo Verde and others 2017). This suggests that the effect of the haptic signal is reduced whereas the visual signal from the deprived eye is boosted under monocular deprivation. Another aspect of multisensory interactions that are affected by monocular deprivation includes the temporal binding of two multisensory stimuli. The deprived eye presents a sharper temporal acuity, whereas the non-deprived eye shows poorer temporal acuity after 1.5 hours of monocular deprivation (Opoku-Baah and Wallace 2020). These results are in line with homeostatic plasticity describing that the brain compensates the deprivation effect by enhancing the signals from the deprived eye.

The effect of monocular deprivation becomes more complex in the long term. One form of this is amblyopia, where visibility in one eye is impaired during development. Amblyopic patients present abnormal spatial vision for the amblyopic eye (Levi 2006). These include impaired visual acuity, contrast-sensitivity function, and spatial distortion. Critically, the impaired spatial vision on the amblyopic eye is found to be associated with abnormal long-range inhibitory interactions (Levi and others 2002; Polat and others 1997). Here, long-range interactions refer to the interactions between a target and flankers. In psychophysics, the visibility of a target stimulus is either enhanced or inhibited by laterally positioned flankers depending on the relative orientation and distance between a target and its flankers (Polat and Sagi 1993). This suggests that the local mechanisms that are tuned to different spatial frequencies and orientations interact with each other. In amblyopic patients, the long-range inhibitory interactions go over longer distances than in normal controls, resulting in increased suppression (Levi and others 2002; Polat and others 1997). Interestingly, the long-range interactions observed in psychophysics are considered to be linked with the long-range horizontal neural connections observed in the visual cortex (Polat and others 1997). These long-range horizontal connections contact other neurons that have similar stimulus preferences, affecting the property of receptive fields (Hirsch and Gilbert 1991). Thus, it was suggested that the amblyopic patients develop abnormal inhibitory spatial interactions via altered long-range horizontal connections and that such change contributes to poor spatial vision (Polat and others 1997). Indeed, when amblyopic patients were trained to reduce the inhibitory spatial interaction on the amblyopic eye, their visual acuity improved (Polat and others 2004).

In addition to impaired spatial vision on the affected eye, amblyopic patients present abnormalities in binocular vision such as impaired binocular fusion (Levi 2006). Abnormalities in binocular vision in amblyopia is linked with altered interocular suppression, which is defined as inhibition of the signals from one eye by the other eye when dissimilar images are presented to each eye. In normal controls, the interocular suppression is comparable between two eyes. However, in amblyopia, the fellow eye has stronger suppression onto the amblyopic eye. When the visual input to the fellow eye is attenuated, amblyopic patients show normal binocular combinations (Huang and others 2009; Mansouri and others 2008). This increased suppression of the fellow eye is supported by neuroimaging measures as well. The BOLD activity in response to stimulation of the amblyopic eye is reduced and delayed compared with the fellow eye (Farivar and others 2011). This effect was observed pronouncedly when the fellow eye was open and presented with a visual stimulus, suggesting that the reduced and delayed BOLD activity in response to stimulation of the amblyopic eye is due to inhibitory effects from the fellow eye.

Impaired monocular and binocular vision in amblyopia can be restored to some extent by visual perceptual learning (Levi and Li 2009). For example, monocular training aimed at reducing long-range inhibitory spatial interactions improve contrast sensitivity and visual acuity for the amblyopic eye (Polat and others 2004). Monocular training that targets reducing interocular suppression also improves visual acuity for the amblyopic eye as well as some binocular functions including binocular fusion (Hess and others 2010; Hess and others 2011; Vedamurthy and others 2015) and eye dominance of the amblyopic eye (Jia and others 2018). Additionally, monocular training on contrast detection at the cutoff spatial frequency in the amblyopic eye was shown to improve binocular function, indicated by increase of binocular steady-state visually evoked potentials (Gu and others 2020).

Other therapeutic strategies that can improve impaired vision in amblyopic adults include occlusion of the amblyopic eye (i.e., inverse occlusion) combined with physical exercise (Lunghi and Sale 2015; Lunghi and others 2019). Inverse occlusion was once used as an alternative therapy for amblyopia with eccentric fixation but was abandoned until recently due to poor effects in children (Malik and others 1970; Von 1965). However, this approach has been reintroduced (Zhou and others 2019), motivated by the finding that monocular deprivation boosts the signals from the deprived eye (Binda and others 2018; Lunghi and others 2011; Lunghi and others 2013). Furthermore, physical exercise has been combined with inverse occlusion because physical exercise enhances plasticity by reducing the release of GABA (Baroncelli and others 2012) or suppressing the activity of GABAergic interneurons in V1 (Kaneko and Stryker 2014). Indeed, amblyopic patients who received the treatment of inverse occlusion with physical exercise presented improved visual acuity (Lunghi and Sale 2015; Lunghi and others 2019) and this improvement lasted for up to 1 year (Lunghi and others 2019).

In animal models, various approaches targeting the removal of molecular brakes are investigated to promote plasticity in the adult visual cortex. These approaches include, but are not limited to, utilizing dark exposure followed by light reintroduction, or manipulating inhibitory signals. Detailed discussion at the molecular level can be found in the following review article (Hensch and Quinlan 2018). Additionally, animal studies have been investigating whether sleep is involved in ocular dominance plasticity induced by monocular deprivation during the critical period. While this topic is beyond the scope of the current review, it is discussed in other papers (Aton and others 2013; Aton and others 2009).

Taken together, the visual cortex boosts deprived eye signals and suppresses non-deprived eye signals when visual input from one eye is blocked. These alterations are likely the result of the homeostatic plasticity that balances the visual system. Such enhanced signals from the deprived eye biases the eye dominance toward the deprived eye and further influences multisensory integration. When the monocular deprivation is prolonged in the long term, both monocular and binocular vision become impaired. In amblyopia, long-range inhibitory interactions become extended, making the spatial vision on the amblyopic eye poor. Furthermore, the fellow eye develops stronger interocular suppression onto the amblyopic eye, leading to poor binocular vision. Critically, impaired monocular and binocular vision can be recovered to some extent via visual training or inverse occlusion combined with physical exercise. In visual perceptual learning, reducing long-range inhibitory interactions or interocular suppression was shown to ameliorate impaired vision. In inverse occlusion, physical exercise appears to boost homeostatic plasticity induced by monocular deprivation.

Binocular Visual Deprivation

Short-term binocular visual deprivation is known to alter cortical excitability. Cortical excitability can be assessed via TMS pulses to the visual cortex to evaluate the minimal intensity required to induce phosphenes—with lower intensities indicating increased cortical excitability. At 45 minutes of blindfolding for sighted participants, visual cortex excitability increased (Boroojerdi and others 2000). This persisted during 3 hours of binocular deprivation, before returning to baseline 2 hours post light restoration. When subjects were blindfolded for 5 consecutive days, cortical excitability initially increased on the first day but then gradually decreased over the 5-day deprivation period to below baseline (Pitskel and others 2007). On blindfold removal, this decreased cortical excitability returned to baseline.

Changes of the cortical excitability induced by short-term visual deprivation appears to be mediated by the excitatory and inhibitory neurotransmission systems. When subjects ingested lorazepam (which increases GABA_A receptor function) or dextromethorphan (an N-methyl-d-aspartate [NMDA] receptor antagonist), or scopolamine (a muscarinic receptor antagonist), the expected cortical excitability increases following binocular deprivation were blocked (Boroojerdi and others 2001).

Changes in cortical excitability may influence concurrent learning. One key study explored the effect of tactile Braille-letter training over 5 days while subjects were either sighted or completely blindfolded. Relative to non-blindfolded controls, this training improved performance, and functionally recruited V1 during Braille tasks in the blindfolded subjects. By day 5 of blindfolding, V1 showed increases in BOLD activity and behavioral performance was susceptible to disruption from TMS to V1 (Merabet and others 2008). One day after blindfold removal, both markers of V1 recruitment were abolished. This suggests the visual cortex can be repurposed for non-visual input when training is performed during visual deprivation, with visual stimulation appearing to prevent any observable markers of this cross-modal recruitment in V1.

The effects of binocular deprivation become more complicated in the long term. For persons with congenital or acquired blindness, the visual cortex becomes responsive to a wide range of non-visual tasks including Braille reading (Burton and others 2002; Kupers and others 2007), echolocation (Voss and others 2006), language (Amedi and others 2003), and mathematics (Amalric and others 2018). Visual cortex recruitment for non-visual tasks appears to be the strongest in the congenitally blind and weaker but significant for those who acquire blindness later in life. This cortical recruitment can occur after training in adulthood, suggesting that sensory regions remain plastic to new computational demands irrespective of their input modality (Heimler and Amedi 2020).

The evidence for visual cortex recruitment for non-visual tasks comes from fMRI, positron emission tomography, and TMS studies. For example, the visual cortex showed enhanced activity during tasks involving Braille reading (Burton and others 2002), auditory localization (Voss and others 2006) as well as sensory substitution devices that convert visual images into either tactile tongue-stimulation (Ptito and others 2005) or sound (Murphy and others 2016; Nau and others 2015) in blind individuals. Interestingly, the sensory substitution task did not recruit the visual cortex during passive listening, but after 10 minutes of training, the task recruited the visual cortex profoundly in blind individuals (Fig. 6A) (Murphy and others 2016). Behavioral performance on these non-visual tasks has been found to be disrupted when TMS was applied to the visual cortex of blind individuals during task performance (Collignon and others 2007; Kupers and others 2007; Merabet and others 2009), suggesting that the visual cortex is recruited for computational processing that is crucial for performing the task.

Figure 6.

(A) Visual cortex is recruited for a vision-to-sound sensory substitution task in congenitally blind and acquired blind individuals after training. The maps represent the blood oxygenation level–dependent (BOLD) activation in response to the vision-to-sound sensory substitution task before and after training. White arrows indicate the visual cortex where greater activation is observed after training compared with before training. (B) Exemplar somatotopic map of tactile sensations in the visual cortex of a blind individual who was trained to use the tongue in a vision-to-touch sensory substitution task. Transcranial magnetic stimulation (TMS) on the visual cortex induced tactile sensations in the tongue. The areas of the tongue sensation are shown in black on the visual cortex where TMS stimulation was applied. (C) Exemplar retinotopic-like map of spatial sound in the visual cortex of a blind expert echolocator. The maps represent the inflated visual cortex of the left (LH) and right hemispheres (RH). The white line indicates the boundary of V1. Each voxel is color-coded to represent the position of sound (in degrees to central vision) with which the voxel’s BOLD activity presents the strongest correlation. Figure A is reproduced from Murphy and others (2016). Figure B is reproduced from Kupers and others (2006) [Copyright (2006) National Academy of Sciences, U.S.A.]. Figure C is reproduced from Norman and Thaler (2019).

Some studies suggest that the visual cortex of congenitally blind individuals retains a topographic organization (Hofstetter and others 2021; Striem-Amit and others 2015). Furthermore, learning spatialized tasks appear to involve the cortical surface in a topographically organized manner. TMS studies revealed that single pulses to different regions of the visual cortex of the blind can induce tactile sensations in different fingertips of Braille readers (Ptito and others 2008) or different areas of the tongue in tongue-based sensory substitution users (Kupers and others 2006) (Fig. 6B). As such, the interface of the skin may become topographically represented on the visual cortex. Further evidence comes from other sensory representations including spatial hearing. Here the spatial sound is mapped topographically in V1, but only for blind expert echolocators (Fig. 6C) (Norman and Thaler 2019). In addition, sound eccentricity is organized comparably to retinotopic eccentricity in sighted people, with the similarity of these maps correlating with echolocation ability.

Cortical recruitment can also extend to higher visual regions—for example, blind and sighted sensory substitution users can utilize the lateral occipital complex for auditory shape processing, a region also used during visual and tactile shape recognition (Amedi and others 2007). Delivering repetitive TMS to this site impaired the object recognition abilities of a blind sensory substitution user (Merabet and others 2009). This shows that plasticity in higher visual cortical regions may operate according to analogous computational demands (i.e., shape) irrespective of input modality (Cecchetti and others 2016).

The processes that underlie the recruitment and repurposing of the visual cortex toward non-visual inputs are underexplored. Theoretical computational models suggest that the recruitment of the visual cortex for processes such as language for the blind is predicted by Hebbian learning during development (Tomasello and others 2019). However, we still lack empirical evidence that reveals what specific neural mechanisms facilitate the observed recruitment of the visual cortex for non-visual tasks as well as how these processes develop over time. For example, it is unknown whether functional repurposing utilizes similar processes to training-induced alterations during visual learning. Future research could focus on how the neural activity emerges from the visual cortex and whether the activity patterns associated with the learned material are reactivated over time in the visual cortex. This can be seen through evaluating the reoccurrence of brain activity patterns that represent previous learning during the following wakefulness and sleep and measuring changes in E/I ratios and spontaneous oscillations including sharp-wave ripple events and spindles. The involvement of homeostatic plasticity can be also examined by evaluating changes in theta and slow wave activities as well as E/I dynamics.

Conclusions

We reviewed the evidence of preserved plasticity in the adult visual cortex through the lenses of visual perceptual learning as well as monocular and binocular visual deprivation. The findings suggest that the adult visual cortex retains plasticity, which enables our brain to be further shaped by experience. In visual perceptual learning, both Hebbian and homeostatic mechanisms act together in the consolidation process. Specifically, the increase of E/I ratio and reactivation of trained visual feature during wakefulness, and the increase of E/I ratio and strengthened spindle activity during NREM sleep favor Hebbian plasticity. On the other hand, the reduction of E/I ratio and increase of theta activity during REM sleep support homeostatic plasticity. While these two mechanisms seem to exert opposite influences on the system, the homeostatic process is thought to preserve the relative synaptic strengths induced by Hebbian plasticity through selective pruning and strengthening of synapses. By contrast, the enhanced neuronal signals from the monocularly deprived eye appears to be an outcome of homeostatic plasticity. Inverse occlusion combined with physical exercise is also thought to enhance homeostatic plasticity triggered by monocular deprivation. Similarly, binocular deprivation appears to boost plasticity, making the visual system available for other computations regardless of input modality. However, the roles of Hebbian and homeostatic plasticity for binocular deprivation are largely unclear and their processes remain to be further elucidated. Identifying these two mechanisms will assist our understanding of how the brain balances plasticity and stability in adulthood. This knowledge may in turn help improve translations to clinical interventions for more effective visual rehabilitation and substitution in individuals with low vision and blindness.