The basic premise that we introduce here is that the circuit architecture of thalamic inputs and the firing mode of thalamic cells may support a mechanism for detecting deviance from cortical predictions. Inputs to the thalamus are classified into drivers, which determine the receptive field properties of TC cells, and modulators, which influence the excitability and functional state of the cells, such as their firing mode (Sherman and Guillery, 1998; Figure 1A). Drivers originate in subcortical afferents from sensory organs or in layer 5 of the cortex. Modulators, on the other hand, originate in several areas, including cortical L6, and contribute most of the synapses on TC cells (Figure 1B). Modulators produce relatively small postsynaptic potentials that are unlikely to trigger spikes but can last hundreds of milliseconds, influencing voltage-and time-dependent channels such as the T-type (“transient”) calcium channels that mediate burst firing (Figure 1C). One of the general functions of modulators is to switch thalamic cells between tonic and burst firing modes (McCormick and von Krosigk, 1992; Bista et al., 2012). Unlike modulator inputs from the brainstem, which present a loose topography and are thought to switch the firing mode of TC cells as a function of general arousal states, like wakefulness and sleep (Briggs and Usrey, 2011; Varela, 2014), the L6 CT inputs are topographically and functionally constrained and could regulate the TC firing mode under context-or stimulus-specific conditions.

The evidence reviewed above suggests that L6 CT neurons inhibit TC cells with dissimilar topography and functional properties and could switch them to burst mode. T-type calcium channels responsible for burst firing are distributed along the dendrites and soma of TC cells (Parajuli et al., 2010; Chen et al., 2017) and can be in one of three states (Figure 1C): inactivated, de-inactivated, and activated. If the membrane potential is hyperpolarized below about −60 mV, the channels enter the de-inactivated state and are primed to become active, thus, “burst-ready.” Subsequent significant membrane depolarization switches T-channels to a transient (40–50 ms) active state, which generates an inward Ca²⁺ current. This current in turn can initiate a burst of somatic Na⁺-K⁺ action potentials with brief inter-spike intervals (~5 ms) between individual action potentials (Lu et al., 1995). Compared to the tonic mode, the burst mode is non-linear and provides increased signal-to-noise properties (Guido et al., 1995; Reinagel et al., 1999). This makes burst firing an efficient mode for indicating non-linear, discrete events, such as sudden changes in the environment or the detection of attended stimuli in the searchlight hypothesis. Furthermore, bursts have higher transmission efficacy at retinogeniculate and TC synapses (Swadlow and Gusev, 2001; Alitto et al., 2019), suggesting that they activate cortex more reliably than tonic spikes (Hu and Agmon, 2016). These properties have led to the proposal that thalamic bursts serve as a discrete attentional signal, or a “wake-up” call to the cortex (Crick, 1984; Reinagel et al., 1999; Sherman, 2001a,b). While this explanation cannot fully account for the function of bursts, given their higher occurrence rates during non-rapid-eye-movement (NREM) sleep (Hirsch et al., 1983; Weyand et al., 2001), it offers a hypothesis for their function during awake behavior.

Bursts have been reported in awake animals during sensory stimulation, suggesting a role in information processing. In general, the proportion of spikes in bursts in TC cells of awake animals is very low, averaging 1–2% (Guido and Weyand, 1995; Weyand et al., 2001; Massaux et al., 2004; Wang et al., 2007; Alitto et al., 2011), but the fraction can be close to 20% in higher order TC cells, those connected with higher order cortical areas (Ramcharan et al., 2005; Varela and Wilson, 2020). However, only these two studies compared burst rates directly between first order (FO, sensory) and higher order (HO) thalamic cells using extracellular electrophysiology, which necessarily relies on the pattern of inter-spike intervals to detect bursts and therefore cannot identify burst spikes directly (Lu et al., 1995; Massaux et al., 2004). In addition, average proportions can be misleading because bursts may occur transiently at high rates during specific behaviors, or in the presence of certain sensory stimuli. For instance, bursts represented 20–30% of the spikes in the somatosensory and auditory thalamus in response to whisker or auditory stimulation (Massaux et al., 2004), although their proportion was low with visual stimulation (Weyand et al., 2001). A key open question is to understand how the variance in the burst rates reported in the literature relates to bottom-up mechanisms, such as input properties, or to top-down processes such as contextual or attentional regulation. In particular, the modulation of TRN activity during covert attention (McAlonan et al., 2006, 2008) could be regulated by L6 activation.

The possibility that bottom-up mechanisms regulate bursts is supported by the finding of bursts after stimuli that are thought to hyperpolarize thalamic cells, such as stimuli with non-preferred features or those that engage the inhibitory surround (Weyand et al., 2001; Lesica and Stanley, 2004; Alitto et al., 2005; Wang et al., 2007). Bursts are also found following periods of sensory quiescence, in the early part of the response to sensory stimulation (Guido et al., 1992; Guido and Weyand, 1995; Alitto et al., 2005; Ortuño et al., 2014; Whitmire et al., 2016). The evidence is more limited regarding the role of CT projections in inducing context-or attention-dependent TC hyperpolarization and bursts. Yet, bursts are observed during behaviors in which the CT system may play a role, such as in dLGN cells after eye saccades (Guido and Weyand, 1995; Martinez-Conde et al., 2002), and in the somatosensory thalamus during whisker twitching, which is thought to be initiated by cortex (Fanselow et al., 2001).

We hypothesized in section 2.1 that the L6 CT projections to TRN can induce inhibition in functionally misaligned TC cells, which predicts that bursts would be more likely in TC cells that represent a mismatch to features (location, stimulus properties, context) represented by active L6 cells. That is, TC cells in “burst-ready” mode may provide a thalamic representation of what is unexpected based on inference from internal cortical models. This mechanism could explain the increase in bursts observed when sensory stimuli are novel compared to after a few presentations under similar attention conditions (Guido and Weyand, 1995; Ortuño et al., 2014). The convergence of L6 CT excitatory and functionally different indirect inhibitory inputs on single TC cells raises the intriguing possibility that a burst-ready predictive code could be implemented at the dendritic level to discriminate inputs (Figure 2), a possibility we assess computationally in Section 3.

Our work builds upon theories that proposed a direct role for the thalamus in top-down covert attention and inference (Crick, 1984; Harth et al., 1987; Mumford, 1991). Our study advances these proposals in several ways (Figure 5). The searchlight hypothesis suggested that covert attention manifests as “rapid bursts of firing” in TRN cells (Figure 5A). However, experiments in the visual system (McAlonan et al., 2008) found a relative decrease in TRN firing in attended compared to non-attended locations in the visual field (although increases in TRN activity were observed with attentional shifts across sensory modalities; McAlonan et al., 2006). In these experiments, animals were well-trained in the task and familiar with the presented visual stimuli, allowing for reliable behavior but possibly seeing less unexpected events. What emerges from the results reviewed here is that the activation of TRN may depend on the occurrence of unexpected events, pointing to the need for experiments where animal expectations are systematically manipulated to investigate their correlation with TRN firing modes. Hierarchical predictive processing theories provide a framework to consider top-down expectations in interpreting thalamic cell responses. These theories suggest that early levels of a hierarchical neural system compute prediction errors (Rao and Ballard, 1999; Friston, 2005). That is, that each level in the circuit hierarchy attempts to predict the responses at previous lower levels via feedback connections, while feedforward connections send the unexplained error to the higher level (Figure 5B). However, these theories have been primarily focused on cortical networks, and it is not clear that the same processes operate in the thalamus, given that thalamic inhibition does not necessarily suppress spiking but instead changes the firing mode. This leads us to propose that CT feedback connections from L6 may specifically modulate the TRN to represent unexpected inputs, and as a result, burst firing may occur in TC cells when they receive an input that deviates from cortical expectations. To gain insight into these possibilities, future experiments on attentional modulation in the thalamus should consider the statistical correlations between stimulus features to understand how they relate to thalamic firing modes. For example, if a vertical line is expected with high probability, a horizontal line would be highly unexpected, and this type of orthogonal stimulus properties may be associated with significantly different membrane potentials and firing mode in TC cells.

Another perspective is the one provided by the “active blackboard” hypothesis (Figure 5C), which sees the thalamus as a region that integrates cortical inferences with sensory inputs to generate the “current best reconstruction of some aspect of the world” (Mumford, 1991). This hypothesis shares similarities with predictive processing theories, as cortical feedback represents predictions. However, it assigns a complex, dynamic role to the thalamus, where its output to downstream brain regions represents an integrated reconstruction of the world. Instead, predictive processing frameworks (Figure 5B) may interpret the thalamus as a circuit that generates prediction errors as a result of comparing its inputs against CT predictions. Based on the observations of burst firing in response to novel stimuli (Section 2.2), we propose that this firing mode may play a role in the predictive computations contributed by thalamic circuits. However, rather than signaling error or novelty, bursts may provide a mechanism to estimate deviance from what was anticipated (Figure 5D). To disentangle these possibilities, new experimental designs could compare thalamic burst rates during the presentation of novel stimuli and of familiar stimuli presented within expected patterns and out of context. By comparing these conditions, we can understand the specific contributions of thalamic firing modes and their relationship to novelty, prediction errors, and deviance estimation.

Along these lines, the hierarchical view of brain systems has dominated the interpretation of neuroscience data for decades, being instrumental in advancing our understanding of the function of the forebrain (Felleman and Van Essen, 1991; Yamins and DiCarlo, 2016). Other views of how information is integrated in brain circuits such as Mumford’s blackboard had received less attention (but see Hawkins et al., 2018). Even von Helmholtz’s unconscious inference ideas, which were theoretical and did not imply hierarchical circuit mechanisms (Meyering, 1989), have evolved into that form under modern predictive processing theories (Keller and Mrsic-Flogel, 2018). Hierarchical circuits have garnered strong experimental support, but additional connectivity patterns are becoming apparent. Non-hierarchical circuits may be particularly important in higher order (HO) thalamic nuclei, those connected with higher order sensory and associative cortical regions (Sherman, 2017). HO TC cells project to multiple cortical areas and receive L6 feedback from them (Giguere and Goldman-Rakic, 1988; Killackey and Sherman, 2003; Llano and Sherman, 2008). We do not know if these different L6 inputs converge on individual cells or form parallel processing streams on different TC cells. HO TC and TRN cells also receive multiple drivers from different hierarchical levels, and these drivers do converge in TC cells, which therefore can integrate information across hierarchical levels (Groh et al., 2014; Sampathkumar et al., 2021; Carroll et al., 2022; Casas-Torremocha et al., 2022). Likewise, projections from one sector of the TRN can diverge to multiple nuclei of the dorsal thalamus (Clemente-Perez et al., 2017; Crabtree, 2018; Li et al., 2020), and modulations of subthreshold membrane potential between TRN sensory modalities have been demonstrated (Kimura, 2014). While hierarchical circuits are beneficial in integrating increasingly complex sensory information (e.g., from detailed features to objects in the visual system; Yamins and DiCarlo, 2016), the connections that bridge across thalamocortical hierarchical levels may facilitate inference at more abstract levels (e.g., concepts, rules).

The proposed hypothesis is constrained by the lack of experimental evidence concerning the involvement of L6 and thalamus in predictive processing. It remains unclear if L6 conveys information on expected and unexpected stimuli or contexts. Nevertheless, the open loops formed by branching L6 CT axons, which both excite and inhibit different sets of thalamic cells, are intriguing within the framework of predictive processing. Perhaps the most basic statistical regularity in the environment, or one of the easiest to infer, arises from orthogonal stimulus properties. For instance, stimuli consistently observed as dark (or vertically oriented, or of high temporal frequency) will be highly unexpected when they present the opposite feature, such as being white (or horizontally oriented, or of low temporal frequency). Having evolved under their constraints, sensory systems reflect the statistics of natural environments (Lewicki, 2002; Li et al., 2003; Freeman et al., 2013; Manookin and Rieke, 2023). Therefore, from an evolutionary point of view, it is reasonable to expect that sensory systems have developed a mechanism to exploit this specific type of statistical regularity through a push-pull system for detecting deviance. The same principle applies to higher-order statistics or cognitive levels, although most higher-level statistics may need to be learned in response to dynamic changes in specific environments.

We used a computational model of a single TC cell with two dendrites to assess basic predictions of the hypothesis in a simplified neuronal architecture. To understand the computations in more complex and realistic circuits, there is a need for experimental data on the convergence and divergence patterns in the thalamocortical circuit at the single cell level. We do not have precise estimates of how many L6 cells converge on individual thalamic cells, or how the response properties of the presynaptic and postsynaptic cells compare to each other.

Topography is a key organizing factor of neuronal response properties both in cortex and thalamus. Anatomical results show that there is a general correspondence between the topographical organization of sensory representations in the cortex and in the thalamus (Deschênes et al., 1998; Rockland, 2021; Shepherd and Yamawaki, 2021). However, it is not completely clear whether the pathways looping between the thalamus and cortex are closed (reciprocal) or open (non-reciprocal) at the single cell level, and with respect to the response properties of the pre-and post-synaptic cells (alignment, misalignment). There is evidence of reciprocal circuits for the direct CT projections to TC cells in several sensory systems (Shepherd and Yamawaki, 2021), but CT axon terminals often spread beyond the thalamic area projecting to L6, suggesting that non-reciprocal loops are present (Rockland, 2019), and the organization at the single-cell level is even less clear regarding the indirect connections via TRN. Also, within a specific cortical location, cell response properties can vary substantially (Sato et al., 2007; Kanold et al., 2014; Rockland, 2019; Wang et al., 2022), which opens the question of how the circuit connections are established with respect to the response properties of pre-and postsynaptic cells, as discussed here. Anatomical and functional data at the single cell level will be key to determine how the alignment and misalignment of functional response properties influence the finer wiring logic and reciprocity of thalamocortical connections.

We do know from electron microscopy studies that TC cells receive about a third of their synapses from L6 inputs (Erişir et al., 1997; Van Horn et al., 2000), and one could imagine a combinatorial code where distinct L6 inputs converge on a thalamic cell (TC or TRN) to regulate its membrane potential depending on the expected likelihood of the combinations of sensory (or more complex) features that the L6 inputs represent. Estimates of L6 convergence onto TC cells suggest that at least 10 L6 cells converge on a TC neuron, up to 100 (Sherman and Koch, 1986). Assuming the minimum of 10 L6 cells, if half of them need to be active to depolarize a TC neuron near the spiking threshold, there are 252 different combinations in which 5 out of 10 cells could be active and represent L6 expectations conveyed to the TC cell in the form of subthreshold depolarization. This mechanism would resemble pattern detection mechanisms proposed in cortical dendrites in which the activation of clusters of NMDA receptors leads to dendritic spikes (Major et al., 2013; Hawkins and Ahmad, 2016). A similar combinatorial code would work on TRN cells, except that the result would be TC hyperpolarization that, according to our proposal, would reflect unexpected feature combinations. In the case of the indirect L6 pathway through TRN, the number of distinct combinations that could be represented as unexpected through subthreshold hyperpolarization would be given by the number of convergent axons from L6 cells onto TRN and from TRN on TC cells, together with the number of active TRN inputs needed to de-inactivate enough T current to produce a burst in a TC neuron. While we lack direct measurements of L6 to TRN convergence, estimates of TRN to TC projections in the visual sector of the TRN (the perigeniculate nucleus, PGN) suggest that the ratio is between10-20 PGN to 1 TC cell, and that each PGN cell produces relatively small IPSPs, less than 2 mV when in tonic mode. Assuming as before that half of the input cells (TRN) need to be active to inhibit the TC sufficiently to de-inactivate the T current, the estimate of potential combinations represented at TRN to TC inputs is similar to those in the L6 to TC projections, except that the indirect pathway would add additional combinations at the L6 to TRN step. Detailed anatomical studies and computational models will be important in figuring out the functional connectivity patterns (convergence, divergence) in the L6-TC circuit and the resulting effect on the membrane potential of individual TC cells.

Another aspect is the spatial distribution of the input synapses on the dendritic tree of TC cells. Data from cortical pyramidal cells demonstrate synaptic architectures that are functionally organized along the dendrites (Sheffield and Dombeck, 2015; Scholl et al., 2017; Kerlin et al., 2019). However, there is limited information regarding the synaptic arrangements on TC dendrites, although there is evidence of dendrite-specific clustering of inhibitory synapses (Crandall and Cox, 2012). Overall, the functional properties of synapses that converge on a given TC dendrite from the L6 direct and indirect inhibitory pathways, and how they relate to the properties of driver synapses are unclear. In the current formulation of the proposed mechanism (Figure 2), we have examined the possibility of dendrite-specific organization with inhibitory synapses that converge on the same dendrite as driver input with similar functional properties (leading to a burst when that driver input is unexpected). However, the mechanism could also be implemented at the cell population level (Figure 4), where individual TC cells (rather than dendrites) may be set in a burst-ready mode based on cortical expectations. In brain slices, the level of depolarization required to evoke a burst by activating individual TC dendrites was substantially greater than that required at the soma (Figure 6 of Connelly et al., 2015). This suggests that burst generation requires multi-dendrite activation of T-channels in vitro, but whether dendrite-specific bursts can be triggered in vivo remains unknown. Ultimately, characterizing the functional and dendritic organization of synapses from different thalamic inputs will be crucial for understanding how individual or ensembles of TC cells integrate inputs, and whether firing mode non-linearities contribute to computing the output that the thalamus relays to downstream brain regions.

The general connectivity of thalamocortical circuits is similar across FO (connected to primary sensory areas) and HO (associative and cognitive) thalamus, with the exception that HO TC cells receive a larger proportion of L6 relative to driver synapses, possibly reflecting the integration of information from more cortical areas in HO cells (Van Horn and Sherman, 2007). Interestingly, there are some differences in burst properties between TRN cells connected to FO and HO TC cells that have implications for our proposal. Intracellular recordings suggest that TRN-FO cells produce bursts more reliably than HO cells. That is, when hyperpolarized to the same level, FO cells are more likely than HO to produce bursts when released from inhibition (Kimura et al., 2012; Clemente-Perez et al., 2017; Li et al., 2020; Martinez-Garcia et al., 2020). This could perhaps reflect the fact that, given that sensed data represents the “ground truth,” a FO circuit can be more “certain” about detecting discrepancies between prediction and observation. Consistent with the TRN findings, HO TC cells in somatosensory and auditory thalamus were less likely to burst when released from inhibition compared to FO TC cells, and the HO bursts had slower latencies (Desai and Varela, 2021). It is possible that HO cells only produce a burst if lower-level models that send them predictive input are very ‘certain’ about their predictions. By being less reliable to burst, HO cells could avoid potential errors in lower-level models. In the language of predictive processing theories, burst intrinsic properties (burst propensity, reliability, and latency) could provide mechanisms to implement “precision,” defined as the confidence associated with predictions (Friston et al., 2014). While the higher reliability to produce bursts in FO cells may seem at odds with the observation of higher proportions of bursts in HO cells recorded in vivo (Ramcharan et al., 2005; Varela and Wilson, 2020), it is important to keep in mind that comparisons of burst reliability are performed by releasing intracellularly recorded cells from a similar hyperpolarization level to assess burst probability. Instead, burst rates in behaving animals may reflect that cells have different hyperpolarization levels, for example because of top-down modulation by L6 (i.e., FO cells may be more reliable to burst, but they may not burst much unless presented with unexpected stimuli).

While our proposal does not require that bursts per se encode specific information about L6 predictions or about the driver input that may trigger a burst, it is possible that burst properties, such as the number of spikes per burst, the latency with which bursts are produced, or the inter-burst interval could provide information about the deviant stimulus or the degree of deviation from L6 predictions. Bursts have in fact been reported to contribute to encoding stimulus features, such as stimulus contrast, transitions from suppressive to preferred stimuli (Reinagel et al., 1999; Alitto et al., 2005; Sanchez et al., 2023), or the degree to which a stimulus matches the preferred properties of the thalamic cell (Martinez-Conde et al., 2002; Massaux et al., 2004).