Contiguity in perception: origins in cellular associative computations

Abstract

Our brains are good at detecting and learning associative structures; according to some linguistic theories, this capacity even constitutes a prerequisite for the development of syntax and compositionality in language and verbalized thought. I will argue that the search for associative motifs in input patterns is an evolutionary old brain function that enables contiguity in sensory perception and orientation in time and space. It has its origins in an elementary material property of cells that is particularly evident at chemical synapses: input-assigned calcium influx that activates calcium sensor proteins involved in memory storage. This machinery for the detection and learning of associative motifs generates knowledge about input relationships and integrates this knowledge into existing networks through updates in connectivity patterns.

Event relationships in perception

David Hume, in the Treatise of Human Nature (1738) [1], developed a theory of thought, in which relationships between sensory percepts (‘Impressions’) in the environment would be reflected (‘copied’) in related trains of thought (‘Ideas’). Hume’s copy principle – which is closely based on Aristotle’s concept of contiguity – provided the foundation for modern associationist theories that claim that our ability to create chains of thoughts can be traced back to the brain’s ability to establish relationships between sensory experiences. Wilhelm Wundt re-introduced Aristotle’s concept of contiguity to the emerging field of psychology. He noted that the fusion (‘Verschmelzung’) of spatial and / or temporal events leads to contiguity in perception and classified the underlying fusion events as a subgroup of associations [2]. Associationist schools of thought in the Psychological Sciences see a wider range of associative brain functions, including aspects of cognition. For example, it has been argued that the capacity of brains to detect causal relationships in natural events relates to the development of syntax in human language [3, 4], a connection that is extended to more simple syntactic structures for communication in non-human primates and birds [5].

In this opinion article, I will argue that – regardless of complexity— all associative brain functions originate from two driving forces: a) the need for creating contiguity (Wundt’s ‘Verschmelzung’) in perception and b) the processing of associative structures as a material property of synapses.

The contiguity argument

Organisms that live in non-static environments need to make sense of associative structures in input patterns [6]. Motion of the environment and/or self-motion will lead to a shift in the perception of orientation points. Knowledge about the relatedness of these orientation points – as defined by their statistical relationships – is crucial for orientation in time and space. If a sensory environment can be described as continuous and non-ending, it is advantageous for an organism to comprehensively perceive it as a contiguous sequence of objects or events as percept interruptions would convey potentially harmful ambiguity in information (i.e. is the environment or the percept incomplete?). The active inference approach developed by Karl Friston and colleagues describes the brain’s attempt to reduce surprise or uncertainty in sensory input as following the free-energy principle, by which such uncertainty is limited by making predictions based on internal models [7]. This motif can also be identified – at perhaps the most elementary level – in the processing of associative input structures.

Associative relationships can be either causal or non-causal, and they can be co-incident or sequential. The latter can be causal, but they can also simply be predictive without being causal. In causal or predictive associations, to borrow terminology from linguistics, an ‘agent’ influences / predicts a ‘patient’ or mathematically:

$$\text{A}=>\text{B }\text{'}\text{If A},\text{ then B}\text{'}$$

Both classical conditioning (learning a relationship between two sequential stimuli) and operant conditioning (learning the relationship between voluntary action and consequence) can be interpreted as the learning of causal associations [8]. From the perspective of an organism moving in the natural world, sequential associative relationships are always predictive, but not always causal. It is contiguity in predictive event relationships that will be discussed here.

Infants as early as 2 months of age show initial knowledge of physical properties of objects, including an appreciation of the effects of gravity on objects in motion [9]. These findings point to the possibility that newborns possess an innate, initial concept of the statistics of object relationships [10]. However, this innate understanding needs to be progressively updated to optimize circuits for behavioral tasks. As humans living in ever changing environments that are oftentimes at least partially human-made (non-natural), we are well aware that innate knowledge of object relationships constantly needs to be expanded and built upon. It often is not possible to clearly distinguish between innate and learned understanding of associative relationships, but it is the non-innate neuronal processing of associative information that will be discussed here. I will also not discuss in detail types of internal predictive models other than associative computations. The reason for this focus is that the detection and learning of associative input structures is the most elementary form of such predictive computations and can be demonstrated at the level of individual cells and its inputs.

Let us approach the problem of contiguity using visual perception as an example. Imagine a situation where your gaze moves slightly. You will experience a smooth transition in perception of the scene. This is true when you are generally familiar with the environment, but to some extent also when you are not. The brain detects associative relationships and learns statistics of the visual environment to be able to store prior expectations [11, 12] and to enable neural computations that reduce the complexity of sensory representation, such as in sparse coding strategies [13]. Making sense of complex visual environments is a difficult task that includes and extends beyond what David Marr introduced as the problem of processing and interpreting structure in the visual world [14], which includes the distinction of objects and background based on Gestalt criteria and their assignment to object categories (both processes involve associative computations). In addition, there is the problem of the relationship of multiple objects in a scene and the interpretation of object relationships when the scene changes. Here, the brain’s ability to give ‘meaning’ to an observation, i.e. the ability to match a scene to prior experiences and expectations, becomes critical. In Figure 1A, the critical role of predicted object relationships is explained using Johannes Vermeer’s painting ‘Girl reading a letter at an open window’ (1659) as an example. Figure 1B shows an example of contiguity violations.

Figure 1: Associative predictions and continuity illusion.

(A) Predicted object relationships in Johannes Vermeer’s ‘Girl reading a letter at an open window’ (1659; Gemäldegalerie Alte Meister, Dresden). If our gaze was initially focused on the girl and our view to the left would be obscured (semi-transparent rectangle), we could still predict that there is an open window on the left side of the scene, because of the direction from which beams of light enter the room and fall onto the girl’s face and clothes, as well as on the wall behind her (arrows). Moreover, we know that in the absence of electric light in those days, she would likely turn towards the window to read a letter. We could predict her reflection in the opened glass window or, if seeing only her reflection, we could anticipate where she stands in the room. These predictions are based on the learning of associative structures (e.g. window – light). (B) Contiguity violations in a Chicago garden scene (photo taken by the author). Violations of expected relationships – such as caused by a mirror arrangement hanging in a tree that causes an unexpected, abrupt change in visible structures and, because of its movement, a sudden blurriness – result in dis-contiguity in perception (arrows pointing to two different directions). Here, the mirrors reflect a different part of a (green) garden scene, leading to a false prediction of contiguity of object 1 (green in the mirror) with object 2 (green outside the mirror). (C) The continuity illusion. The purple line is perceived as continuous and hidden behind the red boxes, although in reality there are three independent purple lines, just as there are three independent grey lines shown for comparison above. It appears that our brains are tuned to create complete percepts that follow ‘Gestalt’ criteria. Contiguity in sensory perception is based on the detection of associative structures. In natural environments, this implies learning about probable proximities between objects.

The problem of contiguity in perception described here is not only faced by humans and perhaps non-human primates. The visual environment changes extremely fast for flying insects. The house fly (Musca domestica), for example, at a top speed of 7.2 km/h [15] covers a distance equivalent to 286 body lengths per second (in comparison, a cheetah reaches 16 body lengths/s [16], and human athletes reach about 6 body lengths/s [17]). During evasive flight maneuvers, flies can reach angular velocities of 5300 °s [18]. Saccadic eye movements in primates can also reach impressive velocities of up to 700–1000°/s, although these are typically followed by fixation times of 200–300ms [19]. These numbers illustrate that perception during movement poses enormous challenges for sensorimotor integration [20]. Specific adaptations enabling fast coupling of visual input (e.g. visually presented threat) to motor responses (evasive flight maneuvers) have been described [18]. The magnitude of this challenge is not generally reduced by low spatial resolution (which might be argued to enhance computation speed). For example, Drosophila fruit flies have developed strategies to enable hyperacute vision by using a microsaccadic image sampling strategy during motion [21]. It appears that predictive knowledge about fast-moving sensory input is required to enable these impressive performances in sensorimotor integration. The ability to predict object relationships in the natural world, and to detect violations in predicted sensory input, requires brains that evolved to handle associative information. Indeed, flies can process relational cues in complex visual patterns and learn associative structures [22]. Bees use floral color information that they associate with olfactory cues when foraging [23]. Innate preferences in color-guided foraging and their adaptation via associative learning have been studied in the Japanese yellow swallowtail butterfly, Papilio xuthus [24]. In experimental tests featuring colored disks on which the butterflies can land, newly emerged butterflies show innate color preferences before training (typically yellow / red). After nectar is repeatedly presented in association with an innately non-preferred color (any color can be used for training), the color preference switches, thus overriding the innate preference (for associative olfactory learning in insects, see [25]). These observations show that the ability to detect and learn causalities and associative relationships is widespread in the animal kingdom, including in invertebrates.

The Search for Contiguity: temporal progression

Contiguity is intimately linked to time. Time progresses, and it does so in a monotonic, continuous (uninterrupted) manner. What seems like a trivial statement, is not, at least not from the perspective of the brain’s representation of time. There is no internal sensor of time. Instead, perception of time relies on the perception of an interval between events and a progression of time between these events [26, 27]. For the detection of associative structures, the order of specific events along a progressive time axis is crucial. In a changing environment during movement, time and space become indicators of event succession [28], which has led to the proposal of hippocampal ‘time cells’ in analogy to ‘place cells’ [29–31]. Similarly, the cerebellum has long been interpreted as a brain structure whose circuit architecture (signals traveling along parallel fibers cross multiple Purkinje cells) is well-suited to encode and generate event sequences [32]. This structure is of particular interest for the representation of temporal relationships in the millisecond range between successive events [26; see also 33]. The role of the cerebellum in forms of associative learning with predictive components underlines the criticality of temporal event relationships.

The brain interprets the environment at the expense of accuracy

The brain needs to compose a comprehensive, contiguous percept of the environment, or at least – as Wilhelm Wundt put it – a reasonable interpretation of it (Wundt used the German term ‘Vorstellung’, which translates better to image / imagination), in a time period that may be too short to compute an accurate percept. The ability to detect and learn associative structures assists in this task, allowing to form critical predictions (note, however, that predictive computations are a general feature of neural circuits that do not always require associative input structures) [34], fill in existing gaps, and quickly group detected objects / events into categories. The latter example – grouping by similarity – is associative in nature (as is the pairing of a specific object with a prototype for each category) and illustrates that associative computations are essential for some cognitive functions as well [35].

This need to interpret and make sense of our environment is also evident in certain types of visual illusions (Fig. 1C). In illusions, relationships are assumed as the most probable outcome – even if incorrect – to create meaning (i.e. a connected line that is visible behind other objects is more likely than a random arrangement of three separate lines). The depiction of the environment in a coherent, meaningful way based on prior expectations is a priority of brain function [36]. It is likely that cognitive processes beyond cellular associative learning mechanisms are involved in these interpretations, but continuity illusions do make the point that the understanding of object relationships in their environment is crucial for organisms.

Biological Mechanisms for Detecting Associative Structures

To begin with, the ability to detect and learn associative input structures is not restricted to nervous systems and their synaptic connections. This ability can be observed in single cells and single-cell organisms. For example, associative memories have been described in gene expression patterns [37]. Associative conditioning – resulting from exposure to electric and chemical stimuli – has been observed in the single-cell organism Amoeba proteus and others [38; see also ref. 39]. For the purpose of the current article, I will focus on associative plasticity capacities of excitatory chemical synapses that, since the discovery of long-term potentiation (LTP) [40] have been linked to learning and memory [e.g. 41] as well as plasticity of sensory receptive fields [e.g. 42].

This selection was made to restrict the discussion to nervous systems, and to associative processes, in which the distinct input patterns that are to be associated are delivered to target cells via separate and independent, activating input pathways. The latter requirement does not apply to plasticity processes at electrical synapses [43] or plasticity of membrane excitability / firing rates [44], which therefore will not be discussed here.

The Hebbian Synapse: a ‘relaxed’ portrayal that enables inclusion of non-cortical plasticity

A rule set for associative learning at chemical (excitatory) synapses was formalized in 1949 by Donald Hebb [45]. The general ability to enhance synaptic strength in an activity-dependent manner – LTP – has later been experimentally observed [40]. A specific prediction from Hebb’s proposal was that excitatory synapses that are consistently contributing to action potential firing in the target neuron – and whose activity therefore precedes the spike firing – will be stabilized / potentiated. A physiological equivalent of this prediction was identified in spike timing-dependent plasticity (STDP). In this phenomenon, the relative timing of presynaptic activity (causing an excitatory postsynaptic potential, EPSP) and the postsynaptic spike matters: EPSPs that precede postsynaptic spikes by only tens of milliseconds cause LTP when this activation pattern is repeatedly applied (the cell ‘interprets’ this activation sequence as the EPSP contributing to spike generation), whereas activation in the reverse order causes long-term depression (LTD) [46–48]. The Na⁺ spike that is evoked near the soma (in the axon initial segment) actively propagates back into the dendrites [49], where it activates voltage-gated calcium conductances, locally enhances spine calcium transients [50], and helps to reach the calcium threshold for LTP induction [51]. Known as the BCM rule, at most types of excitatory synapses calcium transients have to reach a higher threshold for LTP than for LTD induction [52]. Via STDP plasticity, neural networks optimize connectivity patterns, including the development of receptive fields, based on associative motifs in input structure [53]. A contrasting view regarding the classic STDP mechanism has been suggested, which notes that LTP is more sensitive to local dendritic calcium spikes than to backpropagating action potentials [54]. This is an important deviation from strictly applied STDP-type interpretations of the Hebb rule, in which the somatic Na⁺ spike plays a crucial role. However, a more ‘relaxed’ (more general) interpretation is possible as Hebb talked about ‘firing’ of the target cell, which could also include the firing of dendritic calcium spikes. Here, I will therefore use the term ‘Hebbian plasticity’ in a rather inclusive way to describe activity-dependent processes that require some level of postsynaptic spike activity, and cooperative activity of several input synapses (Fig. 2). This wider definition will include STDP-related plasticity rules [see ref. 55] and generalize to calcium-controlled LTP and LTD phenomena [56]. Evidence that STDP indeed conveys predictive information about input correlations across layers of neurons stems from experimental and theoretical work on signal encoding in the salamander retina: it was shown that spiking activity in retinal ganglion cells may encode statistical relationships present in visual information (e.g. in temporal structures) and that – via STDP-based plasticity – feedforward predictions can be read out and learned by downstream neurons [57, 58]. These predictions about the relatedness of stimuli constitute the most basic associative computation that may – after learning – fill input gaps or predict upcoming input components in a temporal or spatial input sequence.

Figure 2: A Hebbian network shows both sequential and coincident associative learning components.

The sequential component enables linear activation sequences of the type A => B, with formal correspondence to linear associative structures in perception and language. The coincident component reflects the need for input cooperativity (A₁ + A₂ => B and B₁ + B₂ => C) to initiate spike firing in the target cells B and C, respectively. The two associativity requirements (sequential and coincident) signify the engagement of a ≥ two-layered network (layer 1: A₁ and A₂, layer 2: B, etc).

Detection and learning of associative relationships in non-cortical brain structures

Hebbian learning has largely been studied in the neocortex and hippocampus. This observation, however, does not exclude the possibility that even strictly interpreted Hebbian-like plasticity mechanisms exist at synapses outside the neocortex or hippocampus, e.g. in the dorsal cochlear nucleus [59], or at cerebellar mossy fibre to granule cell synapses [e.g. 60].

In addition, at types of synapses in non-cortical brain areas Hebbian plasticity has been described that adheres to the more relaxed interpretation of the Hebb rule and thus allows for the demonstration that a plasticity motif with ‘Hebbian’ characteristics mediates associative learning.

Perhaps the best-studied example is supervised plasticity at cerebellar parallel fiber to Purkinje cell synapses. It is worth taking a closer look at this plasticity in detail in the context of contiguity in perception, because the cerebellum plays a role in representing the temporal relationship between successive events [26], and therefore information about associative structures. At these synapses, LTD has a higher calcium threshold than LTP [61–63], and results from co-activation with the climbing fiber input [64]. The term ‘supervised plasticity’ is used, because the polarity of parallel fiber plasticity depends on climbing fiber co-activity. Adult Purkinje cells are classically believed to receive one climbing fiber input, but when multiple primary dendrites exist, two or more climbing fibers may remain connected to a Purkinje cell [65]. Supervised cerebellar plasticity is spike-timing dependent, because climbing fiber activation causes a complex spike in Purkinje cells, but it is not strictly Hebbian [55], because a) the EPSP-spike sequence leads to LTD (‘anti-Hebbian’), and because b) the outcome does not depend on somatic spiking (there is no Na⁺ spike backpropagation in Purkinje cells) [66], but instead it depends on the dendritic complex spike-associated calcium transient [55]. (Note the mechanistic similarity to the STDP mechanism as suggested by Golding et al., ref. 54) In the electrosensory lobe of weakly electric mormyrid fish – a cerebellum-like structure without a climbing fiber input – STDP follows an ‘anti-Hebbian’ rule similar to that found in the cerebellum (LTD is evoked if a dendritic ‘broad spike’ follows EPSP onset within a short time window, while LTP results if the sequence is reversed). Here, the depression component of STDP is seen as enabling a subtraction of learned predictions about sensory input from actual input to make the unpredicted input stand out [67, 68].

Climbing fiber-supervised plasticity at parallel fiber synapses in the mammalian cerebellum – proposed early on by David Marr [69] – is relevant for our consideration of associative learning, because it is one of the synaptic mechanisms enabling classical conditioning, in which causal relationships are learned. Delay eyeblink conditioning (EBC) involves the cerebellum [70]. In EBC, an unconditioned stimulus (US) is presented, which leads to a protective reflex, the closure of the eyelid (unconditioned response, UR). Typically, in animal studies, the US that is selected for this experiment is a brief airpuff that is directed to the cornea. Learning takes place, when a previously neutral stimulus (conditioned stimulus, CS), such as a tone or light signal is repeatedly presented with the US, so that the onset of the CS precedes the US by ~250 ms. After the conditioning period, CS presentation alone will elicit eyelid closure. The animal has learned that the CS is predictive of the potentially harmful US; a causal associative relationship has been established. The paradoxical effect that synaptic depression promotes this form of learning results from the sign-reversal downstream of the Purkinje cell output. These neurons are inhibitory, and therefore a reduction of their excitatory input will lead to a disinhibition of target cells in the cerebellar nuclei, from which further motor commands originate. (Note that the cerebellar output and its plasticity may also control non-motor cognitive functions) [71, 72]. As much as parallel fiber and climbing fiber synapses contribute to LTD at parallel fiber synapses, matching specific roles are assigned to them in eyeblink conditioning. It is well established that the US is conveyed to the cerebellar cortex by climbing fiber activity (the ‘error’ signal), and CS stimuli are ultimately conveyed to Purkinje cells by parallel fiber synapses. The unique circuit architecture of the cerebellum prevents a simple 1:1 comparison of cellular mechanisms to other brain areas, in particular the cerebral cortex, but nevertheless this type of cerebellar learning constitutes the best-characterized type of learning of associative causal relationships thus far.

Sequences of associations can be learned when the conditioned response (CR) developed in operation 1 serves as the CS stimulus in a subsequent learning step 2 [73]. This mechanism may enable long chains of paired associative structures in motor action, language and mental events, complementing mechanisms for sequential learning in cortex and hippocampus [74].

Similarities in associative learning circuits have also been pointed out between the cerebellum and cerebellum-like structures and the insect mushroom body (Fig. 3) [75]. Like cerebellar granule cells, Kenyon cells of the insect mushroom body provide parallel projecting axons that terminate on mushroom body output neurons (MBONs). In honey bees, Kenyon cells make up more than 40% of all neurons in the brain [76] (in Drosophila, it is only 4%), which illustrates a central role in forwarding sensory input at high resolution, similar to cerebellar granule cells, which make up to 80% of neurons in the vertebrate brain [77]. MBON neurons show Purkinje-like dendritic complexity. However, unlike Purkinje cells, which are exclusively GABAergic, only 11 out of 34 MBON neurons per hemisphere in Drosophila use GABA as their transmitter [78]. Modulatory dopaminergic neurons (DANs) provide a second type of synaptic input to MBON neurons. In associative learning, these neurons convey US signals to MBONs [56]; to some extent, they therefore constitute a mushroom body equivalent to the cerebellar climbing fiber input. Mushroom bodies are the insect’s associative learning centers, as has been demonstrated for the learning of olfactory associations [79, 80]. The timing of Kenyon cell input activity to that of the DAN input (via differential activation of type 1 and type 2 dopamine receptors) determines the polarity of synaptic plasticity at Kenyon cell synapses [25], which is an analogy to the climbing fiber controlled plasticity at parallel fiber synapses [61, 62]. It is remarkable that circuit architectures that excel at the detection and learning of associative structures in sensory input can be found in the vertebrate cerebellum and similarly in insects. These are brain circuits that evolved millions of years before the six-layered neocortex developed in mammals.

Figure 3: Similarities in connectivity diagrams between the insect mushroom body and the cerebellum.

MBON cells in the mushroom body assume a role in the circuit comparable to that of cerebellar Purkinje cells. Note, however, that some MBON cells are excitatory and some are inhibitory. Kenyon cells are functional analogues of granule cells, and dopaminergic input fibers show plasticity-supervisory functions resembling those executed by climbing fibers. DAN: dopaminergic neuron; MBON: mushroom body output neuron.

Associative computations as a material property of synapses

All cellular structures have types of co-incidence detectors that principally can detect independent signals in the environment. Over time, these molecular pathways have evolved, and enabled more complex cellular computations. Here, I will focus on known mechanisms for associative plasticity that can be found at chemical synapses. A powerful mechanism is provided by activity-dependent calcium influx and its interplay with calcium sensor proteins, in particular calcium/calmodulin-dependent kinase II (CaMKII). Calcium influx acts as a messenger of neuronal input, shows Hebbian co-activation properties (voltage-dependence), and translates this activity-dependence into biochemical processes of physiological relevance. CaMKII – as a prominent calcium sensor in the postsynaptic density (PSD) and a kinase – acts as a memory molecule [81–83], both at cortical / hippocampal [84] and cerebellar synapses [85, 86]. It controls the phosphorylation level of target proteins such as glutamate receptors, affecting their biophysical properties and membrane insertion state.

Each synaptic input leaves its own calcium signature; these are detected and integrated by CaMKII as a calcium sensor that now becomes the perfect molecular machine for the detection and learning of associative input structures. With a certain spatial and temporal proximity – key parameters in defining associations –calcium signals do interact and may reach the amplitude threshold for activation of CaMKII (or other calcium sensor proteins; Fig. 4). Calcium transients reaching the CaMKII activation threshold promote LTP at cortical / hippocampal synapses and LTD at cerebellar synapses [87]. Notably, the calcium threshold model leaves space for an influence of input activity well beyond the Hebbian range of tens to hundreds of milliseconds, as has been described for the phenomenon of ‘behavioral time scale synaptic plasticity’ [88]. In the cerebellum, low calcium signal amplitudes are sufficient for LTD induction if they are presented for prolonged periods of time [89]. This ‘leaky integrator’ phenomenon can be explained by a shifting calcium threshold under low calcium exposure [62]. Here, the LTD threshold shifts toward a lower calcium amplitude – showing phenomenological correspondence to the threshold shift in the BCM mechanism. It is therefore conceivable that – under specific conditions including spatial proximity – an input signal can be associated with another set of inputs when showing prolonged activity outside the typical Hebbian pairing window.

Figure 4: Calcium sensor proteins detect associative input structures.

Independent excitatory inputs to a target neuron convey information about the presence of objects / patterns in the environment (left). If repeatedly presented in sufficient temporal and spatial proximity – and therefore constituting associative relationships – the input-associated calcium transients will show interactive effects, such as amplitude build-up (right), which allows them to reach the calcium amplitude threshold for the activation of critical calcium sensors, such as CaMKII. As an effector in memory formation, activated CaMKII will facilitate synapse stabilization / potentiation of the respective synaptic inputs, e.g. by enhancing membrane insertion of glutamate receptors. In this way, relationships between input patterns are reflected in discrete, but interacting calcium transients. This read-out of separate inputs to one target cell provides a first step toward the detection of associative object relationships and, ultimately, a cellular substrate of David Hume’s copy principle.

Concluding remarks

Our brains thrive to achieve continuation in what William James called the ‘Stream of Consciousness’ (or ‘Stream of Thought’) that – like an inner monologue – progresses along a monotonic axis of time [90]. The detection and learning of associative structures are at the heart of these processes, constituting a necessity for simple and complex brain functions. These are enabled by elementary, material properties of plastic synapses and even single cells without synapses. This is not to say that a focus on associativity closes the interpretative gap between cellular events in the nervous system and mental representation (see Outstanding Questions); it remains unknown how such signals are ultimately decoded to enable information transfer across the multiscale organization of the brain [91, 92]. However, associativity is a motif that prominently spans multiple scales of brain function; what makes it interesting is that it can be found as a driving force at the evolutionary beginning of nervous systems and, possibly, a driving force for cognitive functions developed in the more recent history of evolution.

Outstanding Questions.

• Is associativity a universal motif in brain function that can be observed across multiple observational scales and functional capacities (e.g. motor coordination)? What mechanisms link associativity across these scales?

• What is the evolutionary origin of associative neural computations? Are single-cell organisms truly able to detect associative input structures? Are instead multi-cellular (neural) networks with chemical synapses required? Did the importance of associativity emerge as a result of organismal mobility (and consequential changes in the environment)?

• Are there non-associative computations that complement associative processes, such as control of cellular / neuronal excitability?

• Are there clinical observations that imply aberrant associative plasticity mechanisms? This possibility is suggested by, for example, the dysfunction of associative long-term synaptic depression in forms of autism spectrum disorder (ASD). Future research will have to investigate whether the central role of associative computations in brain function is reflected in clinical abnormalities upon dysregulation of these computations

Highlights.

• Contiguity in the perception of space and time is a crucial prerequisite for the formation of interpretable sensory experiences and impressions.

• As proposed by Wilhelm Wundt, the ‘fusion’ of spatial and temporal events relies on associative processes (i.e. the detection and learning of associative relationships).

• The role of associations in elementary sensory processes demonstrates the importance of the associativity motif in brain function and has indeed been suggested as a basis of cognitive functions as well.

• Capacities for the processing of associative information are not only present in cortex, but also in non-cortical brain areas, in particular the cerebellum.

• Input-specific calcium transients and calcium sensory proteins linked to plasticity pathways provide essential components of a cell’s associativity detection machinery.