A thalamic bridge from sensory perception to cognition

The ability to adapt to dynamic environments requires tracking multiple signals with variable sensory salience and fluctuating behavioral relevance. This complex process requires integrative crosstalk between sensory and cognitive brain circuits. Functional interactions between cortical and thalamic regions are now considered essential for both sensory perception and cognition but a clear account of the functional link between sensory and cognitive circuits is currently lacking. This review aims to document how thalamic nuclei may effectively act as a bridge allowing to fuse perceptual and cognitive events into meaningful experiences. After highlighting key aspects of thalamocortical circuits such as the classic first-order/higher-order dichotomy, we consider the role of the thalamic reticular nucleus from directed attention to cognition. We next summarize research relying on Pavlovian learning paradigms, showing that both first-order and higher-order thalamic nuclei contribute to associative learning. Finally, we propose that modulator inputs reaching all thalamic nuclei may be critical for integrative purposes when environmental signals are computed. Altogether, the thalamus appears as the bridge linking perception, cognition and possibly affect.


Introduction
Extracting relevant signals from complex environments in order to instruct appropriate behaviors is an essential brain function. The underlying transformation of sensory signals into meaningful actions relies on the coordination between sensory and cognitive brain areas. Classically, cognitive abilities are thought to be largely supported by the cortex and, in particular its prefrontal portion (Donoso et al., 2014). In contrast, subcortical regions, such as thalamic nuclei are believed to relay sensory information to the cortex. Mounting evidence now contradicts this view, showing that a tight interplay between cortical and thalamic nuclei appears to be essential for both sensory and cognition (Halassa and Kastner, 2017;Halassa and Sherman, 2019;Wolff and Vann, 2019), from rodents to humans (Pergola et al., 2018).
Conserved across species and modalities, thalamocortical (TC) systems consist of two principal signaling streams via first-order (FO) and higher-order (HO) thalamic nuclei (Sherman, 2016). The former corresponds to the textbook sensory relay function of the thalamus: FO receives driver inputs capable of directly relaying spiking activity from the periphery. On the contrary HO thalamic nuclei receive driver inputs from layer 5 of the cortex and not from the periphery. These nuclei may thus participate critically to cortical functioning, by implementing a transthalamic route, allowing cortical areas to communicate via parallel cortico-cortical and cortico-thalamo-cortical streams (Sherman, 2016(Sherman, , 2005Sherman and Guillery, 1996). All FO and some HO nuclei are modality specific. For example, auditory signals are relayed via the ventral division of the medial geniculate body (MGv, the FO nucleus) to the primary auditory cortex (A1) while neurons in layer 5 of A1 in turn drive the dorsal division of the medial geniculate body (MGd, the HO nucleus) (Lee and Sherman, 2010). HO are governed by cortically "computed" signals rather than by ascending "raw" sensory signals, which corroborates the idea that HO integrates selected sensory signals from the sensory cortex, that are subsequently fed into recipient cognitive cortical circuits. Thus, in contrast to FO nuclei, who mainly function to relay sensory signals to the cortex, the complex functions of HO thalamic nuclei are to date far from understood.
A current open question in system neuroscience is the specific role played by sensory and cognitive HO thalamic nuclei. It is currently assumed that HO nuclei play highly integrative roles whereas FO nuclei are not expected to achieve specific functions, besides their relay role. However, some classic and more recent data contradict this idea and suggest instead a more intricate functioning of FO and HO nuclei, consistent with the view that the thalamus as a whole could link sensory perception and cognition.
In this review paper, we aim to provide an overview on how thalamic nuclei provide a critical contribution to process environmental signals, from their initial sensory processing to the attribution of a predictive value, which may vary over time. The latter can drastically alter fundamental mechanisms of sensory perception. We will first focus on elementary thalamic mechanisms supporting sensory perception. How environmental signals are perceived is also dependent on attentional factors. We will thus expand on the critical role played by the reticular thalamus on the ability to direct attention toward salient cues. Then we will consider associative learning paradigms that illustrate how thalamic nuclei critically contribute to cognitive processing by assessing the current behavioral relevance of environmental signals. Surprisingly, there is no broad coverage of the large body of data on that topic despite its relevance for the interplay between cognitive and sensory processes. We thus aim to provide a comprehensive review of that literature to fill this gap. Finally, we discuss how thalamic nuclei as a whole may implement neural routes connecting sensory and cognitive signals to fuse them into a meaningful experience and identify some possible research directions to further explore this view. It should be noted that while the primary focus of the review are studies conducted in rodents, primate literature is discussed at various places, especially for conceptual views on how thalamic nuclei are at the crossroad of multiple circuits supporting cognition, emotion and perception.

Main features of thalamocortical circuits from a sensory perspective
Common to all sensory modalities, thalamic activity is determined by two types of synaptic inputs. While driver synapses robustly relay presynaptic input via few but strong giant synapses, modulator synapses alter information transmission via numerous regular sized synapses (Sherman and Guillery, 1998). Electron microscopy and electrophysiological studies in the dorsal thalamus have shown that driver boutons form multiple synaptic contacts with their postsynaptic target (Groh et al., 2008;Guillery, 1966;Hoogland et al., 1991;Viaene et al., 2011) and largely determine the receptive field and activity patterns of their postsynaptic targets. Modulator inputs are characterized by small boutons forming single synapses with the postsynaptic target (Guillery, 1966;Viaene et al., 2011) and can modify the information transmission via both ionotropic and metabotropic glutamate receptors activation. This can therefore impact the signal-to-noise ratio of the driver inputs, for instance by modifying receptive fields (Li et al., 2003;Sherman and Guillery, 1998;Sillito and Jones, 2002;Temereanca and Simons, 2004).
Driver and modulator inputs also differ in aspects of short-term plasticity regarding the postsynaptic response profile. Driver inputs are characterized by a high probability of vesicle release, leading to large amplitude, fast excitatory postsynaptic potential (EPSPs) via ionotropic receptors activation. However, repetitive activation leads to rapid synaptic fatigue, resulting in frequency dependent depression of postsynaptic responses (Mease et al., 2016a;Mo et al., 2017;Sherman and Guillery, 1998;Viaene et al., 2011). Synaptic depression contributes to sensory response adaptation, a phenomenon common to all sensory modalities and well-characterized in FO nuclei such as the ventral posteromedial nucleus (VPM) (Chung et al., 2002;Martin-Cortecero and Nuñez, 2014;Mease et al., 2014). On the other hand, modulators are characterized by a low probability of vesicle release. A single pulse stimulation can thus elicit a small amplitude EPSP, that will increase with repetitive activation in a frequency-dependent manner.
Based on the origin of their driver inputs, thalamic nuclei have been classified as FO and HO nuclei. FO nuclei receive driver input from the periphery and are thus considered to mainly function as thalamic relays for sensory signals en route to the cortex. For example, the VPM receives driver input from trigeminal (V) brainstem nuclear complex and relays whisker information to somatosensory barrel cortex (Feldmeyer, 2012;Rodriguez-Moreno et al., 2020;Viaene et al., 2011). By contrast, HO nuclei are dominated by descending cortico-thalamic (CT) inputs from cortical layer 5 and thus receive sensory signals mainly via the cortex (Mease et al., 2016b) and, to a lesser extent, via ascending sensory pathways . This conserved thalamocortical architecture suggests that sensory signals are relayed via FO nuclei, processed in the cortex and then transmitted via HO to different sensory related cortical areas, thereby implicating HO as a trans-thalamic communication hub between cortical areas (Guillery, 1995;Mo and Sherman, 2019;Sherman, 2016Sherman, , 2005. One of the best characterized HO nuclei is the posterior nucleus (PO), which receives whisker signals predominantly via barrel cortex layer 5B and subsequently relays information back to barrel cortex and to several other cortical areas like primary motor cortex (Casas-Torremocha et al., 2019;Mease et al., 2016b;Mo and Sherman, 2019;Zhang and Bruno, 2019).
While the distinct roles of FO and HO are not entirely understood, the formulation of the transthalamic bridge function of HO nuclei (Sherman, 2016) is also supported by specific cortical innervation patterns. FO nuclei contain specific thalamocortical neurons projecting to small confined zones within one cortical, whereas HO output pathways involve "multi-specific" neurons (Clascá et al., 2016) that broadly innervate several cortical recipient areas (Bosman et al., 2011;Kuramoto et al., 2017b;Pouchelon et al., 2014).
The parallel dual streams of information through both FO and HO nuclei enable not only processing of environmental signals, but also integration with other information through transthalamic communication ( Fig. 1). An important additional feature of thalamocortical circuits is the possibility to fine-tune ascending thalamocortical information as a function of current needs. We thus consider in the following section the prominent role played by the thalamic reticular nucleus (TRN).

The thalamic reticular nucleus as a thalamic searchlight
Due to its anatomical position, its intrinsic physiological properties, and its connectivity, the TRN has long been hypothesized to play a role in attentional control (Crick, 1984;Skinner and Yingling, 1976;Yingling and Skinner, 1976). New converging evidence indicates a fundamental contribution of the TRN to many cognitive processes such as rule switching, cognitive flexibility, fear extinction, spatial navigation and flight behaviors (Dong et al., 2019;Lee et al., 2019;Nakajima and Halassa, 2017;Vantomme et al., 2020;Wimmer et al., 2015). This new research adds to the growing body of literature implicating the TRN in cognitive processing and provides an excellent opportunity to readdress how the TRN may link frontal cortices with somatosensory cortices in the context of attention and cognition. In this section we provide a brief summary of the main physiological features of the TRN before analyzing the literature indicating that the TRN may constitute a thalamic searchlight that is controlled by sensory and frontal cortical areas.

Cortico-reticular gating of incoming sensory information in the thalamus
The ability to direct and maintain attention is an essential ability for survival. Living organisms process numerous sensory signals and distinguishing between relevant versus irrelevant stimuli, which are both computed by the thalamus en route to the sensory cortices (Sherman and Guillery, 2006) is crucial. The TRN has arguably the biggest influence on gating incoming sensory information from the external environment, as it is the primary source of inhibition in FO thalamic nuclei (Guillery, 1995;Halassa and Acsády, 2016;Pinault, 2004). This means that brain regions that project to the TRN can exert influence over thalamic signaling, possibly altering the perceptual salience of the information carried by these signals.
The TRN wraps around the anterolateral thalamus and provides modality specific inhibition to each of the thalamic nuclei (for a review, see Pinault, 2004). It in turn receives driving input from collaterals belonging to the ascending primary thalamic neurons en route to the sensory cortices, and modulatory input from descending cortico-thalamic collaterals Ulrich, 2004, 2003;Sherman and Guillery, 1998). Thus, the TRN can be broadly divided into modality-specific topographic sectors that receive driving inputs from the thalamus, and modulatory inputs from the cortex. Specifically, the dorsocaudal visual sector (visTRN), the ventrocaudal auditory sector (audTRN), and the ventrocentral somatosensory sector (ssTRN) are connected to their respective thalamic nuclei and cortices (Coleman and Mitrofanis, 1996;Shosaku et al., 1984;Shosaku and Sumitomo, 1983). Innervation from sensory cortices originates from descending projections in layer 6 of the cortex that also synapse onto thalamic nuclei of its respective modality (Bourassa et al., 1995;Deschênes and Hu, 1990;Sherman and Guillery, 2006). Descending CT projections vastly outnumber the ascending thalamocortical projections. For example, there are as many as 10 CT projections for every thalamocortical projection in the somatosensory cortex Jones, 2002;Liu et al., 1995). These layer 6 (L6) boutons are excitatory 'modulators 'and exhibit paired-pulse facilitation at both the CT-TC and the CT-TRN synapses (Crandall et al., 2015;Jurgens et al., 2012;Sherman and Guillery, 1998), though the GABAergic TRN-TC synapse shows paired-pulse depression, giving a net output of L6 that can exert either an excitation or inhibition on thalamic relay cells dependent on firing rates (Crandall et al., 2015;Zhang et al., 2017). In sensory cortices, only pyramidal cells in L6 project directly to the TRN but all thalamic nuclei (both FO and HO nuclei) receive this modulatory influence, possibly opening the way for integrative functions (Sherman, 2016). The physiology of this CT influence appears to be complex and highly dynamic with an excitatory-inhibitory balance shifting in an activity-dependent fashion (Crandall et al., 2015;Kirchgessner et al., 2020).
Simple inhibition of TC firing appears however too simple to account for the subtle role of the TRN in highly-evolved functions. As a source of dynamic inhibition to thalamic nuclei, the TRN is in prime position to regulate the availability of the T-Type Ca 2+ current in thalamic cells, and thus whether the thalamic cell is in 'burst' or 'tonic' firing mode (Sherman, 2001). The TRN itself contains at least two different distinctions of cells capable of bursting, and how TRN bursting contributes to cognition is not well understood, though altered TRN bursting is implicated in many neurological disorders; from Schizophrenia to ADHD (Ferrarelli and Tononi, 2017;Krause et al., 2003;Lee et al., 2007;Pratt and Morris, 2015;Wells et al., 2016). Clearly, further work is required to highlight the mechanisms by which the TRN can exert its influence on thalamocortical circuits.

The role of TRN in attention
From rodent to primate, normal attentional functioning is reliant on widespread connectivity between prefrontal areas, FO/HO nuclei and the TRN (McAlonan et al., 2008;Roth et al., 2016;Zhou et al., 2016;Zikopoulos and Barbas, 2007) The specific role of the TRN in attention has been extensively examined in the context of vision. Early studies reported a specific enhanced activation of the visTRN when rats actively explore novel environments (Montero, 1997;Montero et al., 2001). Interestingly, this effect was drastically reduced by lesions restricted to layer 6 of the visual cortex (V1), suggesting an essential role of descending modulatory corticothalamic inputs from sensory areas (Montero, 2000). Similarly, many c-Fos-positive neurons are found in the TRN when rats attend to conditioned but not neutral stimuli, indicating that selective attention may indeed engage TRN activity (McAlonan et al., 2000).
Recent works conducted in rodent provide further support for this view, relying on two-alternative forced choice tasks (2AFC), which typically assess how animals dynamically switch attention from distinct sensory modalities to solve a relatively simple conditional discrimination task. In this task, the mouse is first informed by a single cue whether it should 'Attend to vision' or 'Attend to audition', then it must wait during an anticipatory phase before selecting the correct reward location when presented simultaneously with conflicting auditory and visual stimuli according to the rule for that trial (Wimmer et al., 2015). Recording in opto-tagged visTRN neurons showed increased firing in 'Attend to audition' trials and decreased firing in 'Attend to vision' trials during the anticipatory phase (Wimmer et al., 2015). This is consistent Fig. 1. Schematic drawing of sensory thalamocortical organization. Sensory information from periphery reaches FO nuclei (red box) via driver inputs (blue axons). FO nuclei relay sensory information to cortical layer 4 and TRN (red axons). Subsequently, the cortex relays the processed events to HO nuclei via driver input from layer 5. Finally, HO nuclei relay information to other related cortical areas (green axons). Cortex modulates the activity of FO and HO nuclei via modulator inputs (black axons) and also triggers TRN activity.
with the notion that the TRN implements an attentional filter that may be controlled by top-down signals as a function of current task rule. Using chloride photometry, it was then confirmed that the TRN engages feed-forward inhibition of the lateral geniculate nucleus (LGN), specifically to FO thalamic relay cells that signal to V1.
In a cross-modal attention task assessed in primate, visTRN activity transiently increased when attention shifted from auditory to visual stimuli (McAlonan et al., 2008). Whilst apparently counterintuitive with the before-mentioned gate-keeper role of the TRN, evidence in both rodent and primates suggests that increased firing rates of visTRN units during visual attention appears not to be uncommon, even if the population response tends towards decreased visTRN activity (McAlonan et al., 2008;Wimmer et al., 2015). Subsequent work in primates has demonstrated inverse relationships in firing rates between LGN and visTRN populations during goal-oriented visual attention in single-modal tasks, very similar to the rodent data (Wimmer et al., 2015) as well as an overall decrease in visTRN population responses (McAlonan et al., 2008). Thus, while the link between attention and TRN activity is evident in both rodents and primates, the sign of modulation of individual TRN neurons associated with attentional shifts differs between some studies, possibly as a consequence of procedural or between-species differences.

The TRN as a bridge between sensory perception and cognition
TRN's sectorial segmentation and open-loop organization is well suited to integrate inputs from a wide range of cortical regions that could utilize the TRN for gating information flow through thalamic nuclei with high spatiotemporal precision (Briggs and Usrey, 2008;Pinault and Deschênes, 1998;Willis et al., 2015). The modulation of TRN activity associated with attention shifts has been shown to be controlled by cortical inputs from sensory areas (Montero, 2000).
Recent findings in mice now suggest that the dynamic crosstalk between prefrontal and thalamic regions can suppress distracting information flow to the visual cortex through on-demand recruitment of the TRN by the prelimbic cortex (PL) (Wimmer et al., 2015). Subsequent research has clarified that this PL modulation of TRN activity acts in modality-specific pathways, through prefrontal-striatal projections innervating the caudal globus pallidus (GB), that in turn project to the TRN (Nakajima et al., 2019). Moreover, this prefrontal-striatal pathway also enhances sensory discrimination, possibly by active suppression of distracting information. Interestingly, the activity of the TRN within an extended PL-GB-TRN circuit may become more important when the task is more challenging. In a simple visual Go/No-go task, visTRN firing decreases during stimulus anticipation and presentation but neither its activation nor its inhibition actually affects error rates (Halassa et al., 2014). However, when the cognitive demands of the task increase, i.e. for 2AFC detection tasks, the PL, via the basal ganglia, decreases TRN firing in the sector of the behaviorally relevant stimulus, and increases activity in sectors of the anticipated distractor, which is necessary for successful performance (Nakajima et al., 2019;Wimmer et al., 2015).
In 2AFC tasks, the role of the PL appears to be more important during the anticipatory period whereas interfering with TRN functions is detrimental at any stage of the task. It may be that anticipation, expectation, or even pre-emptive visualization of a learned stimulus may elicit certain activity patterns in the LGN and V1, that are dependent on sustained PL activity (Lundqvist et al., 2016;Schmitt et al., 2017). However, upon stimulus presentation, sensory thalamic signaling and cortico-thalamic feedback could possibly transiently sustain these representations, and the PL may therefore be no longer necessary. Clarifying prefrontal and TRN interactions in other types of task would be very useful to further our understanding of this unique circuit at a broader level.
Prefrontal cortical areas innervate the anterior TRN (Cornwall et al., 1990) via large and small cortical boutons, suggesting that both L6 and L5 may contact these TRN regions Zikopoulos and Barbas, 2006). Whereas innervation of TRN by sensory cortical regions is almost exclusively from L6, the purported innervation from PFC L5 to certain TRN sectors is suggestive of relatively strong driving control of inhibition in HO thalamic signaling. Experimental evidence for prefrontal modulation of sensory HO nuclei via the TRN is limited, yet a few studies have uncovered a possible role for these pathways in cognitive behaviors (Dong et al., 2019;Lee et al., 2019;Vantomme et al., 2020). Activity in the cingulate cortex, that receives input from the orbitofrontal cortex (OFC) and the medial prefrontal cortex (mPFC) among others, can bidirectionally modulate flight behaviors in response to fearful stimuli by inhibiting the interomediodorsal thalamic nucleus via excitation of the limbic TRN (Dong et al., 2019). Modulating the activity of the medial part of the dorsal midline thalamus by rostroventral TRN has been shown to modulate cued fear extinction, through altered signaling to the central amygdala . Finally, presubicular-retrosplenial cortices have been implicated in facilitation of egocentric strategies during spatial navigation through monosynaptic excitation of anterodorsal TRN, resulting in finer tuning of head direction cells in anterior thalamic nuclei (ATN) (Vantomme et al., 2020).
Are reticulo-thalamic interactions solely under cortical control? A recent study uncovered an important role for a poorly known pathway from the basolateral amygdala (BLA) to the TRN for increasing toneevoked responses in the MGv and in the auditory cortex (Aizenberg et al., 2019), possibly opening new research directions for assessing the role of the TRN in associative processes or fearful contexts. The TRN can thus be recruited by a number of brain structures to track the current relevance of environmental stimuli; be it by the PL to flexibly perform cognitive tasks (Nakajima et al., 2019;Wimmer et al., 2015), by sensory association areas through L6-mediated top-down modulation (Ahrens et al., 2015), or even by the amygdala to potentially assign fearful associations (Aizenberg et al., 2019). In addition, a recent study indicated that the TRN may influence both sensory signaling and integrative pathways (Martinez-Garcia et al., 2020), consistent with the idea that this region appears as an important integrative node for cognition and sensation.

The Pavlovian thalamus
Processing environmental signals not only requires to direct attention to salient cues but also to assess their behavioral relevance, especially when they allow to predict important events. Learning episodes during which an initially meaningless sensory cue can acquire a predictive value to signal behaviorally relevant events (i.e. to become a conditioned stimulus, CS) can be readily modelled through the rich theoretical and operational framework of associative learning (Rescorla, 1988). This section aims to analyze the dense literature that has highlighted thalamic contribution to Pavlovian processes. We begin by focusing on the anterior thalamus before considering the role of FO nuclei, embracing both auditory and visual regions to identify functional commonalities beyond modality-specific sensory processing. We then move on to examine the role of HO nuclei, with one example of modality-specific thalamic nucleus and then a summary on the research conducted on both the mediodorsal thalamus (MD) and the reuniens thalamic nuclei (Re), which has tapped into various experimental settings and provided insights on the neural circuits to which these nuclei contribute. Finally we consider the substantial and currently vivid research conducted on thalamic nuclei not currently formally classified as FO and HO, showing that this consideration alone is not sufficient to capture their circuit-level functional contribution.

Anterior thalamic nuclei
The anterior thalamus (technically, anterior thalamic nuclei, ATN) is a unique thalamic region whose fundamental links with memory functions were unambiguously identified in diencephalic patients decades ago (Aggleton and Brown, 1999;Harding et al., 2000). Cognitive impairments following ATN damage are evident in both spatial and nonspatial tasks taxing multiple sensory modalities Law and Smith, 2012;Wolff et al., 2006), indicating that the role of this region is primarily determined by the nature of the cognitive challenge encountered. The anterior thalamic region comprises three distinct nuclei with dissociable properties. For instance, the anteroventral (AV) and the anteromedial (AM) nuclei have been proposed to act together to relay theta oscillations within the Papez circuit while the anterodorsal (AD) nucleus is part of the head-direction cell circuit (Vann and Aggleton, 2004). Analyzing cortical and subcortical afferents has led to the suggestion that the AM could be considered an HO nucleus while both the AV and the AD could be considered FO (Perry and Mitchell, 2019) but there is no definitive conclusion on that topic and further research is needed.
Early evidence showing a specific role of the anterior thalamus in Pavlovian conditioning come from a series of experiments conducted by Gabriel in the rabbit. These studies demonstrated that the AV plays a role in discriminative avoidance learning, during which animals learn that a specific auditory cue signals the occurrence of a mild electric footshock. AV neuronal responses were initially found to reflect learning during conditioning (Gabriel et al., 1977) or reversal of stimulus-outcome contingency (Gabriel et al., 1980) but AV lesions appeared to impair the maintenance or retention of the conditioned avoidance behavior rather than its original acquisition . In rats, acquisition of contextual fear memory is delayed after ATN lesions (Dupire et al., 2013;Marchand et al., 2014). In addition, contextual fear memories formed without the ATN region are less enduring over time (Marchand et al., 2014). Surprisingly, the converse pattern of results was obtained after inactivation of the AD, with impaired recent but not remote contextual fear memory (Lopez et al., 2018). Acquisition of contextual fear memory for predators was found to be impaired by lesioning (Carvalho-Netto et al., 2010) or inactivating the AM (de Lima et al., 2017). Furthermore, latent inhibition was recently shown to be abolished by ATN lesions (Nelson et al., 2018), which fits well with the proposal that the ATN may be important to assess the current relevance of environmental stimuli (Wright et al., 2015).
Collectively, these data indicate that the ATN substantially contribute to Pavlovian processes, especially in the acquisition phase. The effect is clearer when assessing contextual fear memory suggesting that the ATN is more important when there is no particularly salient stimulus to attend. This observation hints toward a role in directing attention to task-relevant stimuli (Wright et al., 2015), independently of their physical attributes. Interestingly, this function has also been suggested for the MD (Wolff et al., 2015a;Wolff and Vann, 2019). Identifying environmental signals that have current behavioral relevance thus appears as a recurring theme for thalamic functions (Wolff and Vann, 2019).

FO nuclei
The auditory thalamus, the medial geniculate body (MGB) has been an important research focus as auditory cues are often used in Pavlovian conditioning procedures. However, only its ventral subdivision corresponds to an actual FO nucleus as both the dorsal and medial subdivisions correspond to HO nuclei (Lee and Sherman, 2010). Interestingly, the frequency receptive fields of MGv neurons appears to be altered by classical conditioning as these neurons are differentially modulated by conditioned and neutral stimuli (Edeline and Weinberger, 1991). This activity may thus reflect aspects of associative learning at a very early stage of stimulus processing. Unfortunately, all subsequent studies aimed at further documenting the role of this region did not differentiate between the ventral and other subdivisions of the MGB (Campeau and Davis, 1995;Heldt and Falls, 1998;LeDoux et al., 1984;Orsini and Maren, 2009) and will therefore be reviewed in the HO nuclei section below.
Lesions of the LGN, the visual thalamus, block the expression of fearpotentiated startle to a visual but not an olfactory conditioned stimulus (Shi and Davis, 2001). Interestingly, electrical stimulation of the LGN can serve as a sufficient conditioned stimulus to elicit eye blink conditioning in the rat (Halverson et al., 2009;Halverson and Freeman, 2010) and reversible LGNv inactivation blocks the expression of visual eye blink conditioning (Steinmetz et al., 2013). LGN neuronal activity habituates during eye blink conditioning, but can recover from habituation when training switches from unpaired to paired conditioning, indicating that the role of the LGN may involve modulation of attention toward behaviorally relevant cues (Kashef et al., 2014). Consistent with this, a majority of LGNd neurons are differentially activated by CS + and CSauditory stimuli, suggesting that arousal-relative associative responding may occur in this visual thalamic nucleus (Cain et al., 2000). Importantly, these experimental data collected in rodents may translate, at least to some extent, to humans. Indeed, a central role for the thalamus was recently demonstrated in humans using a Pavlovian procedure with fearful faces (i.e. visual stimuli) as CS (Lithari et al., 2015).
Altogether, the data reviewed in this section indicate that there are aspects of associative plasticity in FO thalamic nuclei that are necessary for successful conditioning. These data therefore argue against a purely relay function even for these FO nuclei, and suggest instead that they may contribute to cognitive processes. Provisionally, caution is needed when considering the functional implications of the FO/HO dichotomy as at least some FO nuclei appear to also support non-relay functions despite being under the driver influence of the periphery.

HO nuclei
HO nuclei can be roughly divided into two categories, depending on whether they are specific to a sensory modality or not. The medial and the dorsal portions of the MGB are part of the auditory thalamus and will be consider here as the principle example of sensory HO nuclei. The role of the pulvinar (the HO nucleus for vision) will not be considered here as excellent reviews are available (Bourgeois et al., 2020;Halassa and Kastner, 2017;Saalmann and Kastner, 2009). Both the MD and the reuniens thalamic nuclei will be considered as examples of "cognitive" HO nuclei. As they belong to distinct neural circuits, studies focusing on these nuclei provide complementary insights.

Auditory thalamus
Lesions of the MGm do not alter emotional responses to aversive stimuli per se but suppress the conditioned fear responses to auditory but not visual stimuli (LeDoux et al., 1986(LeDoux et al., , 1984, indicative of a specific role in addressing auditory predictive cues. Typically, discriminative fear learning is impaired after MGm lesions (Antunes and Moita, 2010). Recordings in the MGm revealed that the frequency of the receptive fields of these neurons is affected by classical conditioning, suggesting that this could reflect learning about the functional significance of auditory stimuli (Edeline and Weinberger, 1992;Lennartz and Weinberger, 1992). Similar findings were obtained in the rabbit, with MGm lesions blocking differential Pavlovian conditioning of bradychardia (McCabe et al., 1993) while some MGm neurons were shown to respond specifically to conditioned stimuli (Jarrell et al., 1986;Supple and Kapp, 1989). Eye blink conditioning is also prevented by MGm lesions Freeman, 2010, 2006), showing that the MGm conveys critical CS information irrespective of specific task demands. On the contrary, conditioned inhibition is not affected by MGm lesion, indicating that the involvement of the region also depends on the behavioral outcome of the CS and not only on its physical attributes (Heldt and Falls, 1998).
The critical interdependence of connections between the MGm, associative cortex and the amygdala for successful conditioning was established early on using disconnection procedures (Romanski and LeDoux, 1992). For instance, stimulating the MGm produced a robust and long-lasting neuronal response in the lateral nucleus of the amygdala (Clugnet and LeDoux, 1990). MGm plasticity and successful conditioning was then demonstrated to be critically dependent on the activity of the basolateral nucleus of the amygdala (Maren et al., 2003(Maren et al., , 2001. Further research is needed to clarify the nature of the functional interactions at play between the MGm and the amygdala, but it is generally assumed that both the direct thalamo-amygdalar pathway and the indirect thalamo-cortico-amygdalar pathways may support Pavlovian learning (Boatman and Kim, 2006, see also Fig. 2).
Intracellular signaling pathways and protein synthesis in the MGm are necessary for the formation of fear memories (Apergis-Schoute et al., 2005;Parsons et al., 2006). A viral strategy enabling to overexpress CREB in the MGm showed that this manipulation enhanced the formation of auditory fear memory and produced generalization of auditory fear (Han et al., 2008). The MGm is also a known relay for the extinction of a fear memory but may not be, by itself, a critical site of plasticity for its formation (Orsini and Maren, 2009). Finally, its importance for auditory trace conditioning was recently established in the rabbit, showing a role for this region for both the processing of auditory stimuli and the generation of task-related persistent signals (Hoffmann et al., 2018). Collectively, these data indicate that the MGm plays a critical role in auditory fear conditioning and, together with the amygdala, could be considered the root of auditory fear conditioning (Weinberger, 2011) as summarized in Fig. 2.

Cognitive HO nuclei
The MD has been extensively studied, likely due to its strategical connections with the OFC and the BLA in both rats (Alcaraz et al., 2016a;Groenewegen, 1988;Krettek and Price, 1977; but see Matyas et al., 2014 for mice) and primate (Barbas et al., 2011;Timbie and Barbas, 2015) which are key structures for mediating stimulus-outcome associations (Lichtenberg et al., 2017;McDannald et al., 2005). Its potential role in Pavlovian processes was first examined in the rabbit and damage to this region interfere with both differential heart rate conditioning and eye blink responses (Buchanan, 1991(Buchanan, , 1988Buchanan and Thompson, 1990;Powell and Churchwell, 2002). In line with these behavioral data, the activity of MD neurons varies during conditioning and extinction to reflect associative learning Oyoshi et al., 1996). A single study conducted in rats reported that MD lesions reduced contextual fear conditioning (Li et al., 2004). More recently, electrophysiological recordings and alterations of MD activity in behaving mice indicate that the firing mode in the MD modulates the extinction of conditioned stimuli (Lee et al., 2011), thereby supporting the notion that the MD is important for tracking the current relevance of environmental signals. In the context of fear memories, the input from the superior colliculus has also been demonstrated to be important (Baek et al., 2019), showing that a more thorough consideration of thalamic afferents is necessary to understand thalamic integrative functions.
A number of studies have also examined whether the MD controls behavioral choice based on stimulus-outcome associations through appetitive Pavlovian paradigms. While the role of the MD in Pavlovian devaluation tasks appears uncertain (Pickens, 2008), other data point toward a clearer role in using cues to select appropriate actions (Ostlund and Balleine, 2008;Parnaudeau et al., 2015) especially when the integration of Pavlovian and instrumental contingencies is necessary (Alcaraz et al., 2016b;Wolff et al., 2015b). At a circuit level, the function of parallel BLA-to-OFC routes, directly or indirectly through the MD is an outstanding question (Timbie and Barbas, 2015). A direct comparison of BLA versus OFC lesions in Pavlovian procedures indicates that the BLA consistently signals reward while the MD support cue-guided but not outcome-guided responses (Ostlund and Balleine, 2008). Thus a specialization of these pathways is likely with the indirect route through the MD supporting associative processes, possibly including affective import of stimuli (Timbie et al., 2020), while the direct route from the BLA to the OFC may more generally sustain outcome representations. The highly integrative role of the MD may not only arise for amygdalar afferents but also from an input from the ventral pallidum which has been suggested to account for the ability to bias a behavioral response based on predictive cues . Providing a systematic input-output mapping of thalamic neurons would be extremely valuable to unambiguously describe the organization of these circuits. This can now theoretically be achieved at a cellular level relying on the TRIO technique which should stimulate research on that direction in rodent Schwarz et al., 2015). As a whole, the MD may therefore play a critical role in identifying important environmental signals and tracking their behavioral relevance over time (Wolff et al., 2015a). Beyond the downstream inputs mentioned above, this global function is expected to engage intense functional exchanges between the MD and the prefrontal cortex to generate rules categorization transcending sensory modalities (Halassa and Sherman, 2019;Schmitt et al., 2017) or even shape mental representations (Wolff and Vann, 2019).

Fig. 2. A thalamo-amygdalar-cortical circuit
for auditory Pavlovian conditioning. The auditory CS information (orange arrows) reaches the ventral (FO) and the medial (HO) components of the MGB through dual pathways (black arrows indicate existing connections). During conditioning, a tight interplay between the MGm, the BLA and associative cortical areas is likely to be essential (shaded area). While the MGm directly projects to the BLA and cortical areas, the BLA may influence MGm activity through a TRN projection (brown arrows). Overall, the MGm, may connect cortical and temporal areas for integration of auditory information in cognition. Importantly On the other hand, the Re appears as a major thalamic hub to orchestrate functional exchanges between the hippocampus and the prefrontal cortex and has recently attracted considerable attention (Cassel et al., 2013;Cassel and Pereira de Vasconcelos, 2015). Its role in Pavlovian associative processes has, however, been exclusively examined in the context of fear conditioning. The extent to which contextual fear memory can be generalized appears to critically depend on Re activity, suggesting that the Re acts as a gate adjusting the prominence of relevant environmental stimuli that guide behavioral responses (Xu and Sudhof, 2013). Consistent with this study, Re inhibition before conditioning was found to impair contextual fear conditioning freezing whereas Re inhibition before retrieval appeared to increase generalization to the conditioning context (Ramanathan et al., 2018b). Pretraining lesions but not acute chemogenetic inhibition appear to impair contextual fear memory (Quet et al., 2020), which is reminiscent of the effect of intralaminar thalamic lesion (Lopez et al., 2009), possibly due to the similar projections to the prefrontal cortex for these thalamic nuclei. Chronic chemogenetic Re inhibition during consolidation is sufficient to impair remote contextual fear memory (Vetere et al., 2017) showing further an important role of this region in the formation of enduring fear memories. Reconsolidation of fear memory was also suggested to be at least partially supported by the Re (Sierra et al., 2017). Finally, the Re appears to play an important role in mediating the extinction of fear memories (Ramanathan et al., 2018a;Ramanathan and Maren, 2019). Collectively, these data point toward an essential role of this thalamic region in the adaptive regulation of fear conditioned memory. It would be informative to examine whether this region is also involved in appetitive associative learning.

Other thalamic nuclei
At present, many thalamic nuclei cannot be unambiguously classified as FO or HO (Perry and Mitchell, 2019;Varela, 2014). For example, the submedius thalamus is a little known thalamic region reciprocally connected with the OFC (Alcaraz et al., 2016a;Coffield et al., 1992;Yoshida et al., 1992) and is suggested to be either a FO or a HO nucleus depending on a dorsoventral partition (Kuramoto et al., 2017a). This nucleus was nonetheless recently shown to be critical for the updating but not the initial acquisition of stimulus-outcome associations (Alcaraz et al., 2015), which points toward a critical role for the thalamus in adaptive processes and flexible responding. However, it is unclear whether this region is specifically involved in Pavlovian processes per se as its connections with the OFC were also shown to support the updating of action-outcome contingencies (Fresno et al., 2019). Thus, this region could generally support adaptive behaviors, not only those guided by environmental signals.
The paraventricular thalamic nucleus (PVN) is located at the interface between brain rewards circuits (Kirouac, 2015;Millan et al., 2017). PVN neurons innervated by the PFC but also PVN neurons projecting to the nucleus accumbens both respond to predictive cues (Otis et al., 2019(Otis et al., , 2017 suggesting that the PVN could gate associative learning by dynamically representing salient cues (Zhu et al., 2018). Using classic Pavlovian procedures, the retrieval of auditory fear conditioned memory was found to be associated with increased c-Fos expression in the PVN at late time points after learning (Do-Monte et al., 2015). PVN lesions decrease fear expression but do not affect fear acquisition or extinction (Li et al., 2014). The PVN strongly innervates the lateral division of the central amygdala (CeL) and inhibiting CeL-projecting PVN neurons is sufficient to impair both fear conditioning and its expression (Penzo et al., 2015). Conversely, activating these neurons increased conditioned fear expression (Chen and Bi, 2019). A recent proposal posits that the PVN may be important to arbitrate between aversive and appetitive behaviors, when environmental cues have acquired conflicting valences (Choi et al., 2019).
Appetitive conditioning procedures open avenues to examine how predictive environmental signals may themselves acquire motivational value for some but not all individuals. The former are considered 'signtrackers' while the latter are called 'goal-trackers', the difference being that sign-trackers clearly orient behavioral response toward the predictive cue whereas goal-trackers preferentially approach the location of reward delivery (Flagel and Robinson, 2017). The PVN has been suggested to regulate sign-tracking behavior by adjusting the incentive salience of cues (Haight and Flagel, 2014). Consistent with this view, PVN lesions increase sign-tracking while diminishing goal-tracking behaviors (Haight et al., 2015). The PVN could control these behavioral responses by integrating multiple subcortical afferents from the lateral hypothalamus, the medial amygdala and the nucleus accumbens, which are all recruited upon presentation of incentive predictive cues (Haight et al., 2017). Finally, the PVN also receives cortical innervation from the prelimbic cortex and manipulating this prelimbic-PVN pathway differentially affects sign-and goal-trackers, suggesting a cortical top-down control over the process of incentive value attribution (Campus et al., 2019). A topographical organization within the PVN has been suggested to account for its dual role in appetitive (anterior PVN) and aversive (posterior PVN) information processing, suggesting that, overall, this thalamic region may support the ability to use learned information for executing context-appropriate behaviors (Hill-Bowen et al., 2020).
As a whole, data reviewed in this section reveal a complex interplay between amygdalar, thalamic and cortical components to adapt sensory processing to current cognitive demand (see Fig. 2), possibly in relation with emotional processing (John et al., 2016). Interestingly, the thalamic component of this circuit (the MGB and TRN) appears to link together cortical and temporal regions, which is the case for numerous thalamic nuclei as previously noted (Wolff et al., 2015a;Wolff and Vann, 2019). Overall, the thalamus should be considered a major hub in distributed neural circuits, providing supplemental integrative opportunities as it the recipient of multiple subcortical inputs that have not been sufficiently considered. This clearly calls for a critical assessment of these inputs to better understand thalamic contribution within thalamocortical circuits. Table 1 summarizes many of the studies that have documented a role for thalamic nuclei in associative processes underlying Pavlovian conditioning.

Distinctive thalamic features
Based on the previous sections, several key aspects of thalamic functioning are worth highlighting. First, it is striking that the thalamus as a whole continuously processes environmental signals, from their sensory perception to their integration in cognitive constructs. Each aspect of stimuli processing indeed requires a critical thalamic contribution, often with dynamic interplay between cortical and temporal regions. Of interest, this latter point is reminiscent of a pioneer theoretical account positing that the thalamus can be considered a critical functional link between the frontal and the temporal lobes (Warrington and Weiskrantz, 1982). Building on this view, the thalamus may thus bind together information processed in frontal and temporal brain regions (Wolff et al., 2015a) to dynamically adapt the content of mental representations (Wolff and Vann, 2019). Secondly, it is clear that the TRN plays a critical role in controlling signal flows within thalamocortical circuits by adjusting the attentional gain attributed to each signal, taking into account not only its physical salience but also its current behavioral relevance. Often seen as loops (Guo et al., 2017), thalamocortical circuits can thus be dynamically "opened" by the TRN (Zikopoulos and Barbas, 2007). The recent finding that the TRN can be recruited by brain structures such as the BLA (Aizenberg et al., 2019) further underscores the integrated functioning of thalamocortical circuits with temporal structures. Third, the dichotomy between FO and HO nuclei has a clear physiological relevance but caution is warranted in regards to its functional implication. The dynamics of thalamocortical circuit functioning is inherently dependent on an intricate contribution from both types of nuclei and FO nuclei should not be considered less important that HO nuclei in that respect. As summarized in Table 1, various studies indicate that FO nuclei also process impeding signals as a function of prior learning events. Clearly, this is not to be expected from a pure relay station. Perhaps a conceptual reappraisal of the term "relay" in the context of brain circuits is needed. The complex functioning of FO and HO may represent the thalamic bridge that links perception and cognition as we detail in the next section.

The thalamic bridge
Interactions between cortical and thalamic regions appear to play a key role in cognitive functions. To fulfill these functions, cognitive circuits must be informed about sensory events, and indeed several lines of evidence indicate that sensory responses first appear in sensory cortex and spread rostrally to the rostral pole and the dorsal prefrontal cortices (Aronoff et al., 2010). Numerous studies in primates and rodents

HO
MGm LeDoux et al., 1984;Jarrell et al., 1986Supple and Kapp, 1989Clugnet and Ledoux, 1990Edeline and Weinberger, 1992Romanski and LeDoux, 1992McCabe et al., 1993Heldt and Falls, 1998* Maren et al., 2001Apergis-Schoute et al., 2005Halverson and Freeman, 2006Parsons et al., 2006Han et al., 2008Weinberger et al., 2011Hoffman et al., 2018 Orsini and Maren, 2009* *MGm and MGv were both affected MGd Heldt and Falls, 1998MD Li et al., 2004Orona and Gabriel, 1983Buchanan et al., 1988, 1991Buchanan and Thompson, 1990Powell and Churchwell, 2002 Baek observed streams of sensory signals from primary sensory areas to unimodal association areas, multimodal association areas and finally limbic areas (Hoover and Vertes, 2007;Jones and Powell, 1970;Martin-Cortecero and Nuñez, 2016). These sensory streams resemble parallel and sequential processing from sensation to cognition (Barbas, 2015;Mesulam, 1998) with possible relevance for a unified conscious experience (Chanes and Barrett, 2016). The reasons for the existence of parallel cortico-cortical and cortico-thalamocortical routes (Zingg et al., 2014) are not clear to date. Specifically, better understanding how the interactions between each cortical stage, FO and HO thalamic nuclei contribute to the sequential computations of impeding information en route to prefrontal areas is much needed. Based on the distinct properties of driver and modulator inputs, it is tempting to speculate that ascending sensory events relayed by drivers are subsequently transformed into modulator messages on their way to cognitive subnetworks, in order to be integrated with other multimodal messages. Due to the small impact of individual modulator messages the postsynaptic neuron is able to integrate coincident messages from different streams (Fig. 3). Whereas driver messages guarantee robust transmission of single sensory features, modulator inputs allow integrative processes which take many sensory and non-sensory features into account. Indeed, cognitive circuits receive inputs from different sensory modalities (Barbas et al., 2002). Therefore, in order to weight these inputs, their size should be small enough to not overrule other inputs.
Spatial summation along dendrites could explain, how individual small excitatory inputs coming from different sensory circuits, can be integrated in cognitive cortical areas; for example, when a rodent learns to associate an auditory event with a visual event to perform a task. When both signals occur in a short time window, the postsynaptic cell can act as a coincidence detector serving as the structural basis for associative learning and rule detection. The resulting output spikes of PFC neurons may then generate cognitive and behavioral functions; for example, initiating motor activity and increase or decrease attention in a modality-specific manner (Fig. 3). Thus, one would expect to find a larger proportion of modulator inputs in associative areas compared to primary sensory areas. According to this view, interactions between HO and FO nuclei are necessary to achieve behaviorally relevant functions, again pointing toward the need of a critical reappraisal of the 'relay' function in system neuroscience.

Future directions and concluding remarks
Based on the data presented here, we identify several lines of research that may be useful to further our understanding of the contribution of thalamic nuclei in integrative functions. Perhaps the most evident one is highlighted in Table 1: thalamic nuclei which are not currently classified as FO or HO nuclei nonetheless play highly integrative roles. As such, there is a clear need to collect additional data to solve these currently unresolved issues (Perry and Mitchell, 2019;Varela, 2014), with the informed view that this classification alone is not sufficient to identify the functional contribution of these nuclei. The functional role of thalamic nuclei may not be only determined by the origin of their driver inputs. Modulatory inputs from the cortex reach both HO and FO thalamic nuclei and deeply affect the functioning of thalamocortical neurons, particularly when considering branching corticothalamic pathways to the TRN. It is therefore essential to maintain a solid research effort to interrogate these nuclei not only at the level of their physiology, but also at the behavioral level. Second, the data reviewed in this paper suggest a more intricate functioning of FO and HO nuclei in behaviorally relevant functions. Some of these data were collected decades ago and could inspire new research now that more specific tools exist to interrogate thalamic functions in-depth. The activity of HO nuclei may be influenced by prior experience (Edeline and Weinberger, 1991;Shi and Davis, 2001) so determining the inputs that can influence this activity is of major interest. Relying on projection-specific manipulations has already proven to be a successful strategy to reveal some functional features of thalamocortical organization, such as the reciprocity of connections or "re-entry" (Alcaraz et al., 2018;Morceau et al., 2019), and may be useful here. Third, it is Fig. 3. Proposed integration of sensory signals in the cognitive TC system. Output from sensory TC systems, such as auditory and visual (red and green circle respectively) reach cognitive TC system (blue circle) via cortico-cortical pathways (red and green arrows respectively). Multimodal sensory signals converge via weak modulator inputs in cognitive cortical neurons (blue, left panel). Coincident multimodal sensory signals may thus produce cognitive output signals (bottom, blue). clear that the TRN plays a critical and unique role in thalamocortical functioning. Existing data largely confirm Crick's intuition on the TRN acting as a thalamic searchlight (Crick, 1984) and actually extend it beyond directed attention to cognition (Wimmer et al., 2015;Wolff and Vann, 2019). In this context, the recent data showing that the BLA can recruit the TRN to alter the functioning of thalamocortical circuits (Aizenberg et al., 2019) adds weight to the proposal that the BLA-TRN pathway could constitute an emotional gatekeeper that selects signals sent to cortex for further processing in highly emotional situations (John et al., 2016;Zikopoulos and Barbas, 2012). Investigating further whether this connection effectively links emotion and cognition would be of prime value, reminding us that many thalamic regions were once considered as "limbic" (Vogt and Gabriel, 1993).
To conclude, while the data reviewed in this paper are derived mostly from rodent studies, we believe that the core concepts also translate to primates. For example, the role of the BLA-TRN pathway has been examined recently in mice (Aizenberg et al., 2019), but its postulated function builds upon considerations coming from the primate (John et al., 2016;Timbie and Barbas, 2015;Zikopoulos and Barbas, 2012). Similarly, some HO nuclei such as the MD have been extensively examined from rodent to primate, including humans, with a general cross-species agreement when it comes to overall function (Chakraborty et al., 2019;Mitchell, 2015;Mitchell and Chakraborty, 2013;Pergola et al., 2018). Thus, the main functional thalamic features appear to be largely conserved from rodents to primates. While considering subcortical afferents to the thalamus is essential to fully understand thalamic functions, a comprehensive coverage of these pathways would go beyond the scope of this review (see Wolff and Vann, 2019 for more details). Research conducted on primates has provided insights on how cerebellar and dopaminergic innervations impact the functioning of thalamocortical circuits (Barbas et al., 2013;García-Cabezas et al., 2007). A clear perspective for future research is therefore to study these pathways more systematically in rodent models to further our understanding of the contribution of the thalamus in brain functions while pinpointing between-species commonalities and differences (García--Cabezas et al., 2009). Thalamic dysfunction is at the core of numerous mental pathologies such as schizophrenia (Anticevic et al., 2014;Pinault, 2011), obsessive-compulsive disorders (Monteiro and Feng, 2016;Rotge et al., 2012), drug addiction , Down syndrome , Korsakoff syndrome (Harding et al., 2000;Victor et al., 1971) and Alzheimer's disease (Aggleton et al., 2016) Identifying global thalamic functions is therefore valuable as a myriad of pathologies sharing the same issue of long-range disconnectivity may benefit from such advancements (van den Heuvel and Sporns, 2019). Until recently, research on sensory and cognitive aspects of thalamic functioning have been relatively separated. Several lines of research now suggest that a fundamental aspect of thalamic functioning is that it fuses perceptual, emotional and cognitive information into one single meaningful experience. Operationally, this appears to be as close as possible to a definition of consciousness. Decoding the neural bases of consciousness has indeed proven to be a challenging and long journey (Sohn, 2019) which invariably connects with thalamic research (León-Domínguez and León-Carrión, 2019; Llinas et al., 1998;Seth and Baars, 2005).