In the rodent whisker-mediated tactile system, neuronal adaptation is observed across all stages of sensory processing, from the whisker follicle through the brainstem and the thalamus to the somatosensory cortex. Vibrissae, or whiskers are the starting point of this system. The instantaneous velocity of whisker movement is one of the fundamental kinematic features of whisker-mediated sensation in both modes of sensation, the receptive (passive) mode and during whisking (active or generative mode). Whisker velocity has been shown to be related to the radial distance of stationary and moving objects during contact with whiskers (Bagdasarian et al., 2013; Lottem et al., 2015) as well as the speed of the moving object (Lottem et al., 2015). Additionally, profile of whisker velocity determines texture-specific kinetic signatures through sequence of stick-slip events—discrete high-velocity, high-acceleration whisker micro-motions—during contact with objects (Arabzadeh et al., 2005). As such, the body of the literature commonly characterises tactile stimuli in terms of the velocity of whisker movement.

The first stage of sensory processing, the trigeminal ganglion (also known as semilunar ganglion), consists of the cell bodies of pseudo-unipolar neurons with their distal axons arborising the vibrissae follicles and shaft of the mystacial whiskers as mechanosensory receptors (Vincent, 1913; Ma and Woolsey, 1984, for a review see Adibi, 2019). The proximal axons of trigeminal primary sensory neurons innervate the ipsilateral brainstem trigeminal complex (Vincent, 1913; Ma and Woolsey, 1984). Each ganglion cell innervates only one whisker follicle (Fitzgerald, 1940; Zucker and Welker, 1969; Dykes, 1975; Gibson and Welker, 1983; Rice et al., 1986; Lichtenstein et al., 1990). Each follicle is innervated by 150–200 myelinated and around 100 unmyelinated distal axons of trigeminal ganglion neurons (Lee and Woolsey, 1975; Waite and Cragg, 1982; Renehan and Munger, 1986; Rice et al., 1986, 1997; Henderson and Jacquin, 1995). The nerve terminals and mechanoreceptors are of various types, morphologies and distributions (Melaragno and Montagna, 1953) such as Merkel cell-neurite complexes, lanceolate receptors, club-like endings, Ruffini-like corpuscles—also referred to as reticular ending—and free nerve endings (Renehan and Munger, 1986; Rice et al., 1986; Ebara et al., 2002). Early studies classified the trigeminal ganglion neurons into slowly adapting and rapidly adapting based on their response profile within the first milliseconds to a rapid change in whisker angle. Rapidly adapting receptors do not elicit response to maintained whisker deflection, while slowly adapting receptors respond to sustained whisker deflection (Talbot et al., 1968; Kwegyir-Afful et al., 2008). The relative time interval between velocity-independent first spike latency of rapidly adapting neurons and velocity-dependent first spike latency of slowly adapting neurons accurately and reliably encodes whisker movement velocity (Lottem et al., 2015). Whisker velocity is also encoded, although less robustly, by the firing rates of slowly adapting neurons (Shoykhet et al., 2000; Bale et al., 2013; Lottem et al., 2015). Merkel endings are the most prominent mechanoreceptors with slowly adapting characteristics (type I skin Merkel endings in mouse trunk and ring-sinus Merkel receptors in rat vibrissal follicles) (Iggo and Muir, 1969; Woodbury and Koerber, 2007; Furuta et al., 2020). Club-like, rete-ridge collar Merkel and lanceolate receptors are rapidly adapting while reticular-like type-I Ruffini endings exhibit slowly adapting characteristics (Li and Ginty, 2014; Tonomura et al., 2015; Furuta et al., 2020). Rapidly adapting ganglion cells have generally higher velocity thresholds (Zucker and Welker, 1969; Lichtenstein et al., 1990). The time course of adaptation in trigeminal ganglion neurons and mechanoreceptors is in the order of a few tens of milliseconds. In contrast to cortical (Khatri et al., 2004; Musall et al., 2014; Allitt et al., 2017; Kheradpezhouh et al., 2017, see Figure 1), thalamic (Hartings et al., 2003; Khatri et al., 2004; Ganmor et al., 2010) and brainstem neurons (Mohar et al., 2013), the response of trigeminal ganglion neurons to repeated deflections at stimulation frequencies as high as 18 Hz exhibits little adaptation (Ganmor et al., 2010).

Proximal axons of the first-order trigeminal ganglion neurons innervate the two sensory nuclei in ipsilateral brainstem trigeminal complex: the principal sensory nucleus (PrV) and the spinal nucleus (SpV). The PrV and SpV interpolaris sub-nucleus (SpVi) provide the main sensory input to the thalamus forming the starting point of the two major parallel streams of somatosensory signals: the lemniscal and the paralemniscal pathways (Yu et al., 2006, for a review see Adibi, 2019). Neurons in these two sub-nuclei exhibit opposite intensity-dependent adaptation profiles to repeated deflections of the principal whisker (Figure 2); PrV neurons adapt less to higher intensity stimuli (Figures 2A,B), while neurons in SpVi exhibit increased adaptation as the intensity of deflections increases (Mohar et al., 2013, see Figures 2E,F). The intensity-dependent adaptation feature in PrV neurons is preserved at the level of VPM and cortical neurons (Ganmor et al., 2010; Mohar et al., 2013, 2015). While the neuronal mechanisms of intensity-dependent adaptation in SpVi and PrV neurons remain unknown, these findings suggest that neuronal adaptation in PrV may be due to the inter-subnuclear inhibition of PrV neurons by the SpVi (Furuta et al., 2008); as the intensity of stimulus increases, SpVi neurons adapt more causing a greater disinhibition of the PrV neurons. The disinhibition of PrV neurons to repeated whisker deflections, in turn, decreases the level of adaptation at high-intensity stimulation regime; PrV neurons adapt less to higher intensity stimuli, while neurons in SpVi exhibit increased adaptation as the intensity of deflections increases (Mohar et al., 2013). The intensity-dependent pattern of adaptation in PrV neurons further is preserved at the level of VPM and cortical neurons (Ganmor et al., 2010; Mohar et al., 2013, 2015). Increasing the amplitude and velocity of whisker deflections does not increase the adaptation of synaptic responses in layer 4 neurons of the somatosensory cortex, but rather entails less adaptation (Ganmor et al., 2010). Importantly, previous studies (Timofeeva et al., 2004) suggested that inter-subnuclear interactions between SpVi inputs shape the receptive field size of PrV neurons. Indeed, the pattern of intensity-dependent adaptation in PrV neurons is reversed when the adjacent whisker is stimulated. That is, increasing the intensity of stimulation entails less adaptation when stimulation intensity increases.

Although the two trigeminal nuclei encode the intensity of whisker stimuli in a similar manner under non-adapted condition (Figures 2C,G), adaptation introduces distinct changes in the coding behaviour of these two nuclei (Mohar et al., 2015). Under adaptation, PrV neurons better encode the fluctuations of the stimulus at high intensity regimes (Figures 2C,D), whereas SpVi neurons better encode weak tactile stimuli (Figures 2G,H). A similar pattern was also observed at the level of the subthreshold synaptic potentials (Mohar et al., 2015). As the neuronal adaptation linearises the response function of PrV neurons at high-intensity stimulation regime, it linearises the response function of SpVi neurons at low-intensity stimulation regime. Thus, the two parallel routes of stimulus intensity processing through PrV and SpVi together enhance the overall coding of stimulus intensity for a broader stimulation range. These findings suggest that neurons belonging to these two brainstem trigeminal nuclei may encode the intensity of stimulation together in a context-dependent manner untangling the coding ambiguity associated with response adaptation in different stimulation contexts.

Along the pathway from periphery to cortex, adaptation exhibits stronger effect on neuronal responses. Peripheral trigeminal ganglion neurons exhibit less adaptation than neurons in the principal nucleus of the brainstem trigeminal complex (PrV). Sensory thalamic neurons in the VPM, in turn, exhibit a higher level of adaptation than neurons in trigeminal complex, and less adaptation compared to cortical neurons (Khatri et al., 2004; Ganmor et al., 2010, also see Figures 3C,D). As mentioned earlier, the degree of adaptation depends on the frequency of whisker stimulation (Ahissar et al., 2000; Khatri et al., 2004; Heiss et al., 2008; Kheradpezhouh et al., 2017; Latimer et al., 2019), the amplitude and velocity of deflections (Ganmor et al., 2010; Adibi et al., 2013b; Mohar et al., 2013), and the cortical state (Castro-Alamancos, 2004a; Katz et al., 2012). The response of the neurons decreases to consecutive individual whisker deflections with an exponential decay (Figure 1C). As the stimulation frequency increases, neurons adapt stronger and at a faster rate (Khatri et al., 2004; Kheradpezhouh et al., 2017, also see Figure 1D). A significant subset of cortical neurons exhibit response facilitation over the course of repetitive stimulation at frequencies of 4–10 Hz (Brecht and Sakmann, 2002; Garabedian et al., 2003; Derdikman et al., 2006; Kheradpezhouh et al., 2017, see Figure 1E). The facilitation is then followed by a reduced responsiveness to further subsequent deflections (Kheradpezhouh et al., 2017). Similar facilitation effects are reported in response to repetitive optogenetic excitation of layer 6 corticothalamic neurons where the evoked response in layer 5a pyramidal neurons as well as fast-spiking inter-neurons in both layer 4 and 5a increased due to activation of facilitating synapses (Kim et al., 2014). Prolonged optogenetic activation of layer 6 corticothalamic neurons resulted in a hyper-polarisation of VPM neurons followed by depolarisation, shifting the mode of sensory responses from bursting to single-spike (Mease et al., 2014). Subsequently, VPM neurons exhibit reduced adaptation to 8 Hz repetitive whisker deflections during prolonged optogenetic stimulation of layer 6 neurons. This suggests corticothalamic feedback shapes both the gain and the temporal profile of sensory processing in cortex by controlling the gating of sensory information in VPM. Post-adaptation response facilitation was observed to stimulation at a few hundred milliseconds after the adapting repetitive stimulation in approximately a third of cortical neurons (Malina et al., 2013). Thalamic neurons, on the other hand, do not exhibit post-adaptation response facilitation revealing that facilitation does not emerge from thalamic neurons. Recordings at different holding currents revealed this facilitation is a result of a faster recovery of excitation compared to inhibition. Adaptation also is shown to decrease cross-whisker suppression, similar to the reduction in surround suppression in visual system (Higley and Contreras, 2006; Ramirez et al., 2014).

Early electrophysiology studies revealed that on average, neuronal responses from layers 2/3 and 5a exhibit stronger adaptation than cortical layers 4 and 5b (main lemniscal input layers) neurons (Ahissar et al., 2001; Derdikman et al., 2006). These findings are consistent with stronger adaptation in the posterior medial (POm) thalamus than in VPM (Diamond et al., 1992; Sosnik et al., 2001). In a detailed study of adaptation to sequences of stick-slip events across cortical laminae, Allitt et al. (2017) found stronger adaptation in supra-granular layers compared to layers 4 and 5. Lower layer 3 showed rates of adaptation that lie between that of layers 2/upper 3 and layers 4/5. Despite strong responses to high-speed protraction (at 654°/s) resulting in suppressed responses to the initial stick-slip event in the sequence, across all laminae, the rate of adaptation vs. frequency did not change with the speed of protraction (654°/s vs. 32.7°/s). However, layer 2 neurons represent initial texture-defining stick-slip events with temporal fidelity and relatively high firing rates irrespective of protraction speed, and only exhibit adaptation to the subsequent repetitive stick-slip stimuli, and not to the strong evoked responses by fast whisker protraction. Interestingly, stick-slip stimuli in Allitt et al. (2017) drive weaker rates of adaptation at stimulation frequencies as high as 34 Hz compared to pulsatile whisker deflections at similar range of frequencies which disrupt texture encoding reported in previous studies (Ahissar et al., 2001; Chung et al., 2002; Khatri et al., 2004).

In the somatosensory cortex, both response latency and response peak time (to repeated whisker deflections) show a consistent increase over time with consecutive stimulations, irrespective of the direction and amount of the change in response rate (facilitation and suppression, as shown in Figure 1G, Allitt et al., 2017; Kheradpezhouh et al., 2017). This delayed response trend was also observed for neurons exhibiting little adaptation as well as those neurons exhibiting either decreased or increased responsiveness with repetitive stimulation. At stimulation frequencies >30 Hz, a large increase in latency from the 1st to 2nd stimulus was observed, followed by a decrease in response latency to a relative steady-state latency longer than that to the 1st stimulus and shorter than that for the 2nd stimulus (Allitt et al., 2017). These findings indicate that the context-dependent changes in the evoked response latency is governed by additional mechanisms than those underlying decrease or increase in the responsiveness.

As shown in Figure 2, adaptation exhibits differential effects depending on the intensity, frequency and potentially other physical parameters of sensory stimulation (also see Katz et al., 2006; Ganmor et al., 2010; Lampl and Katz, 2017; Katz and Lampl, 2021). Thus, it is important to characterise adaptation effects across a range of stimulus intensities. Using multi-electrode extracellular electrophysiology in anaesthetised rats (Figure 3A), Adibi et al. (2013b) quantified how neuronal adaptation modifies the input-output response function of neurons as a function of the whisker deflection intensity (amplitude). The neuronal input-output functions (also known as neurometric functions) typically exhibit an increase in the mean neuronal responses with stimulus amplitude (Figures 3B–D), along with decreased trial-by-trial variability (in terms of Fano factor, Figure 3E). Adibi et al. (2013b) delivered sustained sinusoidal vibrations at various amplitudes to whiskers to induce different levels of sensory adaptation, and then quantified the neuronal input-output function under each level of adaptation (Figures 3A,B). The findings revealed adaptation induces a rightward shift in the neuronal characteristic response functions (both the mean and variability against stimulus amplitude, Figure 3). The magnitude of the shift in neuronal responses depends on the magnitude of adapting stimulus; adaptation shifted the threshold of neuronal responses (the lowest stimulus intensity to which the evoked response is significantly higher than baseline activity) to stimulation amplitudes above that of adapting stimulus (Figures 3C,D). While adaptation shifts the neuronal characteristic functions (response rate and variability), it maintains the relationship between the two across different adaptation states. This was confirmed by a regression analysis between the response variability and mean (Adibi et al., 2013b). Thus, adaptation transfers the operating point of neurons to lower rates with higher variability. It is worth to note that the lateral shift in response function lowers overall responsiveness (spike counts averaged across the whole stimulus range) which in turn, suggests a lower metabolic cost. Consequently, the coding accuracy (in terms of Fisher Information) peaks at amplitudes above the adapting stimulus (Figure 3F, see also Adibi et al., 2013a). Sensory adaptation produces systematic rightward shifts in the stimulus region with elevated coding efficiency consistent with the shift in the evoked neuronal response thresholds to amplitudes higher than that of the adapting stimulus (Adibi et al., 2013b).

Shared neural variability is a ubiquitous phenomenon in neural networks. Conventionally, neuronal activity has been characterised by the average and variance of responses over multiple trials. However, single neuron statistics do not capture the stochastic and dynamic characteristics of a neural network. Correlated fluctuations across neurons—known as noise correlations—are one of the central bases of the recent theories of neural computation (Pouget et al., 2013). These correlations are shown to affect the information content of population activity in the cortex (Averbeck et al., 2006; Adibi et al., 2013a,b). Previous electrophysiology studies indicated sensory stimulation decorrelates the population responses in somatosensory cortex (Middleton et al., 2012; Adibi et al., 2013b). Similar stimulus-driven decorrelation was observed in the primary visual cortex of primates (Kohn and Smith, 2005), middle temporal (MT) cortex (Ponce-Alvarez et al., 2013) and anterior superior temporal sulcus of macaque monkeys (Oram, 2011). Sensory adaptation shifts the profile of noise correlations in cortex along the stimulus amplitude axis similar to the shift in the other response statistics including mean and variability of responses (Figures 3, 4). These parallel shifts result in maintaining the relationship between noise correlations and the mean firing rate across different states of adaptation (Adibi et al., 2013b). Thus, the net effect of sensory adaptation in the cortex is to decrease the overall neuronal response rate across the stimuli while increasing the total variability as well as correlations in variability (noise correlation) across neurons in the cortex.

In contrast, at the upstream sensory stage to somatosensory cortex, in VPM, while thalamic neurons exhibit adaptation in form of a shift in the neurometric curves (Figure 3C), the noise correlation does not show visible difference across stimuli and adaptation states (Figure 4A). This was the case, even though the firing rate of thalamic neurons, on average, increased over two-fold from about 30 Hz at spontaneous level to 60 Hz with increasing the stimulus amplitude. In contrast, on average, a less than 25 Hz change in the response of cortical neurons to stimulation was accompanied by a halved level of noise correlations in the somatosensory cortex (Figure 4B).

By extracellular electrophysiology using a 10 × 10 electrode array from the somatosensory cortex (Figures 5A,B), Sabri et al. (2016) showed that neurons have a specific sequence of activation with respect to the population which is anatomically organised (Figure 5C). Additionally, the strength of pairwise correlations in the ongoing spontaneous activity of neuronal clusters decreases with the distance between the electrodes (Figure 5D), consistent with similar findings in other cortical areas (Smith and Kohn, 2008; Rothschild et al., 2010; Solomon et al., 2015). Correlations, on average, are stronger for units that better encode the sensory stimuli (Figure 5D), and predict the correlations in the evoked response fluctuations to sensory stimuli (i.e., noise correlations, Figure 5E) as well as signal correlations (see Sabri et al., 2016). The strength of correlations in spontaneous state (i.e., non-adapted) and adapted state (during sustained sensory stimulation) were highly correlated (Figures 5F,G, also see Sabri et al., 2016), indicating that neuronal adaptation maintains the spatiotemporal dynamics of population activity within the cortical networks; the functional connectivity map based on these correlations resembles the two-dimensional anatomical organisation of electrode locations (Figures 5H,I), and maintains its organisation across states of adaptation. Similarly, our unpublished data indicates neuronal adaptation maintains the spatial organisation of synchrony in the cortex as captured by the phase coupling of field potential oscillations (data not shown).

The shifts in the neuronal response functions by sensory adaptation cause an increased signal correlation (Figure 5J); sensory adaptation reduces the level of network heterogeneity by shifting the response function of neurons, aligning their responsive range with respect to the adaptor. Similar homeostatic effect of adaptation has been reported in the primary visual cortex of anaesthetised cats (Benucci et al., 2013). Benucci et al. (2013) observed that adaptation to a given orientation maintained the equality in responsiveness and the independence in orientation selectivity across the population.

How do the adaptation-induced shifts in characteristic functions (mean, variability and noise correlations) affect the efficiency of readout mechanisms of neuronal activity in downstream areas? A biologically plausible and efficient yet simple readout mechanism of the neuronal responses is a linear combination of the neuronal responses in a downstream neuron (decoder, see Figure 6A). The coefficients of the linear combination identify the synaptic weights between the neurons, and may be optimised to maximise the flow of sensory information or the decoder's discrimination performance (Figure 6B). The optimum weights depend on the amount of information each individual upstream neuron carries about the sensory stimuli (Figure 6C) and the structure of response co-variabilities across the population of downstream neurons (Adibi et al., 2014). This optimal linear readout determines an upper boundary of coding efficiency using the linear integration framework. Adibi et al. (2014) optimised the readout in two manners: (i) the pairwise-optimal readout scheme where for any pair of stimuli, the linear combination weights were optimised to maximise discriminability. And (ii) the groupwise-optimal readout scheme where an identical set of weights were optimised to maximise the discrimination across all stimuli. The discriminability of neuronal responses under pairwise-optimal readout provides an upper bound for the performance of the groupwise-optimal readout. The groupwise-optimal readout approaches its upper bound when the neuronal responses to sensory stimuli are linearly correlated. This is equivalent to a maximal level of signal correlation in the population responses. In order to apply the appropriate set of weights, the pairwise-optimal readout scheme requires a priori knowledge about the pair of stimuli to be discriminated. Thus, the groupwise-optimal readout is arguably a more biologically plausible scheme.

Adibi et al. (2014) found that for either readout scheme, adaptation enhances the discriminability for stimuli higher in amplitude than the adaptor, while there is a decline in discriminability if both stimuli are lower than the adaptor (Figure 6D). The magnitude of the effect was larger for the groupwise-optimal readout compared to the pairwise-optimal readout. These findings represent a shift in discriminability from low amplitudes to amplitudes higher than the adapting stimulus, consistent with the observed rightward shift in the neuronal response functions in Figure 3D. Additionally, adaptation increases the number of stimulus pairs with enhanced discriminability based on population responses (shades of red in Figure 6D).

As a result of the shifts in neuronal responses (Figure 3), the optimal readout weights which maximise discriminability between a pair of stimuli in the non-adapted state are expected to maximise discriminability between a new pair of stimuli that are in effect simply shifted by the adapting stimulus intensity. This predicts a high level of generalisation of the optimal readout across different states of adaptation. Adibi et al. (2014) verified this by quantifying the discriminability obtained from a non-adaptive readout—which its weights were optimised in the non-adapted state—relative to an adaptive readout—which its weights were optimised under the adaptation state. The results revealed that on average, the non-adaptive readout discriminability was 97 and 90% of that of the adaptive readout, for the 6 μm and 12 μm adaptation states, respectively (Figure 6E), indicating a markedly high level of generalisation of the optimal readout across different states of adaptation.

Noise correlations have been shown to depend on the stimulus features such as intensity-dependence in the somatosensory system (Adibi et al., 2013b) as well as in other sensory systems (Kohn and Smith, 2005; Oram, 2011; Middleton et al., 2012; Ponce-Alvarez et al., 2013). This suggests that instantaneous or short-term correlation structures may potentially provide an additional channel of information for sensory processing. The optimal linear readout scheme provides a framework for studying the effect of noise correlations on the efficiency of sensory processing in different adaptation conditions. Adibi et al. (2014) showed that neuronal adaptation increases signal (see Figure 5J) and noise correlations in population responses. By increasing signal and noise correlations, adaptation increases the redundancy of population responses. This adaptation-induced redundancy, can potentially limit the capacity of the cortical network to encode sensory information. The increased redundancy, in turn, may enhance the accuracy with which population responses represent sensory stimuli on trial-by-trial basis. The effect of noise correlation on information encoding/decoding depends on the direction of noise correlation (in the multi-dimensional space of joint population activity) relative to the direction of signal (Averbeck et al., 2006; Adibi et al., 2013b). In the non-adapted condition, noise correlations improve the accuracy of encoding/decoding for some populations and in some other populations, they were detrimental to population coding (Adibi et al., 2014). Similar opposing effects of noise correlation were observed in awake animals performing a texture discrimination task (Safaai et al., 2013). The differential effects of noise correlation can be attributed to the heterogeneity of neuronal populations; in a heterogeneous population, different neurons may exhibit a variety of signal directions in their responses relative to noise. This leads to opposing effects of noise correlation in the non-adapted state. In the adapted state, however, with decreased level of heterogeneity (Figure 5J), the population responses are more homogenised, showing less diversity in their direction of signal relative to noise. This leads to entirely detrimental effect of noise correlations in the adapted states (Adibi et al., 2014). Compatible with this scenario, it has been observed that noise correlations under sensory adaptation were always detrimental to information encoding/decoding (Adibi et al., 2014). The magnitude of the effect of noise correlations was greater in adapted states than non-adapted condition.

Based on these results, one might predict that ignoring noise correlations would be more detrimental to the performance of the readout under adaptation. On the contrary, ignoring noise correlations in the readout by taking into account only the diagonal elements of the pairwise neuronal response covariance matrix (denoted by subscript “diag” in Figure 6F) was less detrimental under adaptation compared to the non-adapted state. This discrepancy can be explained in terms of a greater increase in signal correlations relative to noise correlations under adaptation (Adibi et al., 2014).

Fisher information (Fisher, 1922) is a well-known measure of coding accuracy that quantifies the amount of information that the neuronal responses carry about the sensory stimulus upon which the distribution of the neuronal responses depends. This measure has been used to characterise the effect of neuronal adaptation on the efficiency of coding in visual, auditory and somatosensory systems (Dean et al., 2005; Durant et al., 2007; Gutnisky and Dragoi, 2008; Adibi et al., 2013a). Adaptation can be considered as the procedure of matching the neuronal responses based on the distribution of the stimulus—a procedure known as ‘equalisation' (Laughlin, 1981; Nadal and Parga, 1994)—in order to maintain the efficiency (or optimality) of neuronal code. For the biological case where the neuronal response variance is stimulus dependent (Churchland et al., 2010; Adibi et al., 2013b), the optimality will be obtained when the square root of the Fisher information function is equal to the distribution of the stimulus. Thus, the peak of Fisher information function should be aligned with the most frequent stimulus in the environment. This is compatible with the Linsker's infomax principle (Linsker, 1988; van Hateren, 1992) and is equivalent to Barlow's redundancy reduction principle (Barlow, 1961, 2001; Atick, 1992; Redlich, 1993).

In vision, neuronal adaptation is shown to maintain the efficiency of the neuronal responses by scaling the response function of neurons with changes in the variance of input (Brenner et al., 2000), or by shifting the neuronal response functions with changes in the mean of input distribution. These adjustments have been reported in contrast adaptation in visual system (Ohzawa et al., 1982, 1985; Sclar et al., 1989; Solomon et al., 2004) and in sound level adaptation in auditory system (Dean et al., 2005) resulting in enhanced accuracy around the adapting (most frequent) stimulus (Dean et al., 2005; Durant et al., 2007). In somatosensory cortex, however, the shift in the coding accuracy to stimulus amplitudes above the adapting stimulus (Figure 3F) is not exactly consistent with the notion of equalisation or information maximisation. Equalisation predicts the peak of Fisher information aligns with the most frequent stimulus (here adapting stimulus). However, the Fisher information profile peaks at above adapting stimulus intensities (Figure 3F). As a result, neuronal adaptation filters out the most frequent features of the stimulus, and in turn, aligns the most sensitive portion of response curve (equivalent to peak of the Fisher information) to amplitudes above adapting stimulus. This potentially reflects a balance between homeostatic regulation of metabolic energy costs of spikes and maintenance of information content of neuronal responses, and in turn, tunes the network to the critical point of deviant detection. In the realm of whisker-mediated tactile world, for rodents, the most frequent stimuli may constitute the ambient noise in the surrounding environment, while behaviourally significant stimuli are those above the most frequent stimulus region. Adaptation can be considered as a mechanism to shift or scale the neuronal functions to maintain their accuracy for significant stimulus region by equalisation with respect to a “significance” function instead of the stimulus probability distribution function. In visual and auditory systems, however, adaptation tunes the neuronal responses to maintain the acuity for frequent stimuli. The more frequent stimuli, in these modalities, may be considered to constitute functionally or behaviourally significant stimuli.

The transformation of physical attributes of the sensory environment into spiking activity is analogous to the concept of an “encoder” in the framework of coding theory in the realm of communications. In this framework, the neuronal computation is analogous to either that of the “source coding” or “channel coding.” An efficient source code is the one that maximises its entropy according to the distribution of sensory input. It is equivalent to minimising the redundancy in the neural code. Channel coding, on the other hand, adds patterns of redundancy in the transmitted signal to reduce the decoding error rate at the receiver over transmission through an erroneous channel. In an information theoretic framework, Adibi et al. (2013a,b) showed the amount of information in the neuronal responses about the amplitude of the sensory stimuli (in terms of the mutual information between the neuronal response and stimulus amplitude) increases with adaptation, hence enhancing coding efficiency (Figure 7A). This leads to the prediction that the sensory cortex may act as an adaptive entropy maximiser that increases the entropy of its codes (spike rate) (Attneave, 1954; Barlow, 1961; Srinivasan et al., 1982; Atick, 1992) similar to an optimum source coder in the realm of communications (Shannon, 1948). To identify the source of adaptation-induced enhancement in coding efficiency, Adibi et al. (2013a) decomposed the information content of neuronal responses (in terms of mutual information between neuronal responses and sensory stimuli) into its two fundamental components: the entropy of neuronal responses and the conditional response entropy given stimulus. Adaptation decreases the response entropy (Figure 7B) and the conditional response entropy (Figure 7C) at both the level of single neurons and the pooled activity of neuronal populations. The net effect of adaptation is to increase the mutual information between stimulus and neuronal responses. The information transmitted by a single spike also increases under adaptation, even when the overall rate of activity is matched across non-adapted and adaptation states (see Figure 7D).

The adaptation-induced increase in the information content of neuronal responses can be explained by a higher number of stimulus pairs for which adaptation increases Shannon information (and also discriminability) than the number of stimulus pairs (at amplitudes lower than adapting stimulus) of which adaptation reduces Shannon information (and discriminability) (Figure 6D, also see Adibi et al., 2013b). This lead to the net increase in the mutual information between neuronal responses and sensory stimuli and is consistent with the rightward shift in the Fisher information (Figure 3F).

Sensory adaptation desensitises the tactile sensory system during exposure to sustained or continuous stimulation. After some time, we tend to not notice ongoing sensory stimulation such as the scratching of a shirt on our body. At the perceptual level, the perceived intensity of tactile stimuli exponentially decreases over time during adaptation (Berglund and Berglund, 1970). The response of sensory neurons also exhibit similar exponential reduction trend at multiple stages of sensory processing (Hartings et al., 2003; Khatri et al., 2004; Musall et al., 2014; Allitt et al., 2017; Kheradpezhouh et al., 2017; Lampl and Katz, 2017). Consistently, as mentioned earlier, sensory stimuli with lower intensity than that of the adapting stimulus evoke little neuronal responses in the somatosensory cortex and thalamus (Adibi et al., 2013b, see also Figure 3). This reduced neuronal responsiveness to prevailing sensory stimuli, hence, provides a noise reduction mechanism to filter ambient stimuli at neuronal and cognitive levels.

The fundamental basis of neural communication and brain function is through action potentials that neurons generate in order to transfer information to other neurons. The cost of a single action potential is high, with a net cost of ~2.4 × 10⁹ ATP molecules per action potential (Lennie, 2003). The fraction of energy consumption in neocortex associated with neural signalling is estimated to be 52% of total energy expenditure (Attwell and Laughlin, 2001; Lennie, 2003). This severely limits the number of action potentials that neuronal populations may persistently generate in response to sustained or repeated sensory stimuli in the environment. In the somatosensory cortex, adaptation improves neural coding efficiency at a reduced metabolic cost associated with spiking, due to a net decrease in neuronal responses (Adibi et al., 2013b, also see Figures 1B, 3D). Similarly, in the primary visual cortex, adaptation equalises population responses to stimulus orientations with different statistics, maintaining the overall rate of spiking averaged over time (Benucci et al., 2013).

Survival in a dynamic and changing environment requires animals to detect unexpected sensory cues that signal necessary commodities (e.g., food and water), mates or danger from ambient sensory stimuli. For an urban rat searching for food in a dimly lit street, the tactile vibrations travelling through the asphalt from the movement of an approaching dog or car provide accurate estimate of the distance from the impending danger. Namib Desert golden moles process seismic sensory signals to detect prey (Narins et al., 1997). An animal, hence, should efficiently and quickly identify salient stimuli from the continuous stream of sensory signals in changing environments and react with appropriate behavioural responses. In tactile, behaviourally relevant events are potentially those with a higher intensity/acceleration compared with the prevailing sensory stimuli. This is particularly true for the generative mode of the whisker-mediated sensory system where interactions with textures and objects produce changes in the whisker trajectory beyond the baseline whisking action (Adibi, 2019). Sensory adaptation serves as a neuronal mechanism for salient stimulus detection by adjusting the sensitive region of the neuronal response functions to stimulus intensities above the background level (Adibi et al., 2013b; Musall et al., 2014, also see Figures 3 and 9).

In the natural environment, the prevailing diet of sensory stimuli varies over time. An efficient neural code adaptively matches the limited discriminative range of neuronal responses according to the distribution of sensory stimuli (Barlow and Földiák, 1989). This constitutes scaling the discriminative region of neuronal responses according to the variance of sensory stimuli (variance adaptation), as well as shifting the most discriminative point of the neuronal responses to the most frequent sensory stimulus (mean adaptation). Previous studies observed that variance adaptation maintains the information content of neuronal responses (Maravall et al., 2007, 2013). As for the mean adaptation, at the neuronal level, adaptation enhances discriminability of neuronal responses to stimuli around the adapting stimulus (Figure 6D, also see Adibi et al., 2014). We predict that the enhanced neuronal discrimination improves perceptual discrimination performances. Similar improved perceptual effects have been observed in humans for vibrotactile amplitude (Goble and Hollins, 1993) and frequency (Goble and Hollins, 1994) discrimination. In animal literature, improved performance was observed in rats performing a spatial whisker discrimination task (Figures 10E–H, also see Ollerenshaw et al., 2014). Future studies are required to investigate the effect of adaptation on perceptual discriminability.

During repetitive stimulation, adaptation leads to higher neuronal responsiveness at some frequencies than other frequencies, and therefore exhibiting frequency-filtering properties. Using different frequencies of whisker stimulation, Kheradpezhouh et al. (2017) observed that the net evoked response of over 90% of cortical neurons was at stimulation frequencies other than the maximum frequency (see also Figure 1). This constitutes a low- and band-pass frequency filtering property of sensory adaptation. However, thalamic neurons show more high-pass frequency response properties to sinusoidal stimulation at frequencies up to 40 Hz (Hartings et al., 2003). This could arise from lower spiking responses to sinusoidal stimuli at lower frequencies and hence at lower mean speed (Arabzadeh et al., 2003). A small yet significant subset of cortical neurons exhibit response facilitation (see Figure 1E) which is equivalent to high-pass filtering. Consistent with the diversity among cortical neurons in their frequency response profile (Allitt et al., 2017; Kheradpezhouh et al., 2017), at synaptic level, also diverse filtering properties have been identified across neurons (Anwar et al., 2017). During synaptic plasticity, facilitating synapses serve as high-pass filters as they are stronger at high pre-synaptic spiking frequencies, while depressing synapses serve as low-pass filters as they are stronger at low pre-synaptic spiking frequencies.

A misconception in the literature is to consider adaptation as a high-pass filter due to its slow adaptive process (for instance see Benda, 2021). Based on this view, high-frequency stimulus components that change on time scales faster than the adaptation processes are transmitted with a higher gain than lower frequency components. The flaw in this interpretation is that a slow filter in the time domain—which is characterised by a long impulse response—is assumed to be a low-pass filter in the frequency domain. However, there are no direct links between the time domain characterisation of a filter (slow or fast) and its frequency-domain characterisations (in terms of the passband frequency range of the filter). In fact, the frequency passband of a slow filter (with long impulse response function) and a fast filter (with short impulse response function) can be around any frequency. Hence, irrespective of the length of impulse response function (slow vs. fast), a filter may exhibit various low-pass, band-pass or high-pass properties.

Neurons conventionally are considered as coincidence detector or integrators depending on the time interval over which they accumulate and integrate input spikes (Abeles, 1982; König et al., 1996; Kisley and Gerstein, 1999). These modes of operation determine the way neural networks encode information; as rate coding scheme, or temporal coding scheme. Neuronal adaptation can be considered as a potential mechanism to shift the operating mode of the neuronal networks in a continuum between the two extreme modes of coincidence detector and temporal integrator. This can be implicated at the circuit level through different degrees of suppressive adaptation in either integrator or coincidence detector neurons within a neuronal population, or at single cell level through mechanisms of short term synaptic plasticity (D́ıaz-Quesada et al., 2014; Anwar et al., 2017) or through subthreshold adaptation modulation of the slope of the membrane potential (Kisley and Gerstein, 1999). Cortical layer 4 neurons including spiny stellate neurons which receive major thalamic input in barrel cortex have short integration intervals of a few milliseconds (Egger et al., 1999; Bruno, 2011; Adibi, 2019) and are driven by weak but synchronous thalamocortical input (Roy and Alloway, 2001; Bruno and Sakmann, 2006; Wang et al., 2010a,b). Varani et al. (2021) recently found that selective optogenetic inhibition of layer 4 neurons decreases sub-threshold responses to whisker deflections in the preferred direction of layer 2/3 neurons, while it increases responses to deflections in the non-preferred direction, leading to a change in the direction tuning. This allows a broader integration of signals in these neurons. During adaptation, thalamo-cortical synapses in layer 4 exhibit short-term depression (Chung et al., 2002; Lundstrom et al., 2010; D́ıaz-Quesada et al., 2014), potentially moving the operation point of cortical circuits from coincidence detection of thalamic inputs toward integration of inter-laminar and cortico-cortical inputs in layer 2/3 or infragranular neurons (for instance, see Jordan and Keller, 2020). Consistently, our unpublished data indicates that circumventing the sensory input to layer 4 neurons through direct optogenetic stimulation of layer 2/3 pyramidal neurons results in little adaptation in neuronal responses across cortical layers in the primary somatosensory cortex of mice.

Previous electrophysiological studies revealed that neurons in the primary somatosensory cortex of rats encode vibrotactile stimuli in terms of the mean speed of whisker movement (Arabzadeh et al., 2003, 2004). The mean speed is equivalent to the product of the two fundamental features of vibrotactile stimuli: frequency and amplitude. While an increase in either feature increases the activity of cortical neurons, no measure of neuronal response (firing rates or temporal patterns) explicitly encodes one principal feature independently of another. This representation forms the basis of whisker-mediated tactile sensation in awake rats as well (Adibi et al., 2012). Consequently, two distinct stimuli with identical products of their frequency and amplitude are indistinguishable based on cortical neuronal responses as well as at the perceptual level. Stimulus dependent properties of neuronal adaptation (Ganmor et al., 2010; Adibi et al., 2013b; Mohar et al., 2013, 2015) potentially provides a neuronal mechanism to disambiguate encoding of the principal features of tactile stimuli from one another. Increases in the frequency and amplitude of stimulation has differential effects on the adaptation profile of neuronal responses. Thalamic and cortical neurons exhibit a higher level of response depression with increases in the frequency of stimulation (Kheradpezhouh et al., 2017). On the contrary, increased amplitude of stimulation results in lower level of adaptation (Ganmor et al., 2010; Mohar et al., 2013, 2015).

Adaptation alters the representation of external stimuli in a context-dependent manner, introducing response ambiguity. That is, an identical stimulus may evoke different responses depending on the context of stimulation, or conversely, different stimuli may evoke an identical response as context shifts. Under adaptation, PrV neurons better encode the fluctuations of the stimulus intensity when the intensity of stimuli is high, whereas neurons in the SpVi better encode weak tactile stimuli (Mohar et al., 2013, 2015). Together, the two nuclei provide and improve the overall coding of stimulus intensity at different stimulation intensity regimes (Mohar et al., 2015). The differential stimulus-dependent adaptation properties in the two parallel pathways of tactile system, the lemniscal and paralemniscal pathways, hence, may help in reducing the inherent ambiguity of neural coding of stimulus features in different adaptation conditions.

Majority of cortical neurons across different layers of whisker area of the somatosensory cortex exhibit multi-whisker receptive fields (Simons, 1978; Armstrong-James and Fox, 1987; Moore and Nelson, 1998; Ghazanfar and Nicolelis, 1999; Brecht and Sakmann, 2002; Brecht et al., 2003). Various spatial properties of receptive fields including the principal whisker, size, response latency and centre of mass in majority of cortical neurons exhibit stimulus-dependent changes (Le Cam et al., 2011). Furthermore, feature encoding properties of cortical neurons changes with the level of spatial correlation in multi-whisker sensory stimuli (Estebanez et al., 2012). Analogously, temporal multi-whisker stimulation patterns (see Figure 8) through whisker-specific adaptation mechanisms in cortical neurons (Katz et al., 2006; Ramirez et al., 2014) can potentially adjust the receptive field properties of neurons during different modes of behaviour in the environment; while sustained stimulation of principal whisker reduces the responsiveness to that specific whisker, this adaptation does not transfer to adjacent whiskers (Katz et al., 2006). This specificity, thus, maintains the responsiveness to stimulation of adjacent whiskers. Consequently, in the adapted state, stimulation of adjacent whiskers evokes a higher level of response relative to adapted responses to principal whisker stimulation. These results in a broader spatial extent of functional receptive field compared to the non-adapted state. Consistently, Ramirez et al. (2014) showed that while surround inputs in the non-adapted state are suppressive, as previously reported in the literature (Simons, 1985; Simons and Carvell, 1989; Brumberg et al., 1996), under adaptation, they are facilitatory and enhance the evoked responses to deflections of the corresponding principal whisker at the level of sub-threshold (Figures 8B–D) and spiking activity (Figure 8F). The adaptation-induced facilitation was stronger in layer 4, 5, and 6 neurons (Figure 8G). Conversely, adaptation by stimulation of all whiskers causes narrowing of the receptive fields (Katz et al., 2006; Ramirez et al., 2014). Adaptation also has been shown to affect multi-whisker integration of tactile stimuli (Figure 8E); integration of unadapted sub-threshold neuronal responses (in terms of PSPs) and spiking activity to preferred multi-whisker stimuli were highly sub-linear (Mirabella et al., 2001; Ramirez et al., 2014). However, when presented in a background of multi-whisker stimulation (adapted state), multi-whisker integration of responses were more linear (Ramirez et al., 2014). Consistently, Ego-Stengel et al. (2005) found that when principal and adjacent whiskers were simulated together at 0.5Hz, 59% of cortical neurons exhibit significant suppressive interactions, whereas response facilitation was found in only 6% of neurons. In contrast, at 8 Hz, a significant supra-linear summation was observed in 19% of the cells, with stronger effect along an arc compared to along a row. Such dynamic changes in the receptive fields could be a potential neuronal basis of invariant coding in whisker system. For instance, it might help to locate the position of whisker contact with respect to the head or the body instead of whiskers (for instance see Curtis and Kleinfeld, 2009). This position invariant information can potentially give rise to whisker-mediated coordination, and contribute to spatio-topic representations such as those in grid cells in the entorhinal cortex (Hafting et al., 2005) or head-direction cells in classic Papez circuit (Taube, 1998). Further experiments in awake and anaesthetised animals are required to understand adaptive changes in the multi-whisker receptive field of cortical neurons and their functional role. The adaptive adjustment of receptive fields forms the spatial dual of the adaptive shift between integration and coincidence detection modes in the time domain.