Hierarchical feature binding in a spiking neural network model of the primate ventral visual pathway Authors: Brian Gardner, Patrick T. McCarthy, Joseph Chrol-Cannon, Dan F. M. Goodman, Simon R. Schultz, Giovanni Lo Iacono, Simon M. Stringer, Abigail Morrison, Daniele Marinazzo, Abigail Morrison, Daniele Marinazzo, Abigail Morrison, Daniele Marinazzo, Abigail Morrison, Daniele Marinazzo Published: 2026 ABSTRACT Feature binding - how the brain encodes which features are part of other features to form representations of the coherent objects we perceive - remains an unsolved problem in neuroscience. Despite progress towards a solution, major theories either lack detailed explanations at the neuronal level or rely on biologically unrealistic simplifications, and none adequately account for the representation of hierarchical information, which is crucial to our perception of the world. To address this, a solution termed binding by polychrony has been proposed to explain how hierarchical feature relationships may be encoded at the neuronal level in a biologically realistic system. This theory relies on a phenomenon known as polychronization, where groups of neurons fire in precisely coordinated, time-locked sequences, leading to the emergence of regularly repeating spatiotemporal patterns that might encode these relationships. In this study, we explore binding by polychrony through simulations of a spiking neural network that closely aligns with the structural organisation of the primate ventral visual pathway, incorporating bottom-up, top-down, and lateral synaptic connections. By exposing the network to collections of related 2D object shapes from ecologically realistic datasets and applying spike-timing-dependent plasticity, the network self-organises such that individual neurons respond selectively to specific shape features. Furthermore, the network exhibits polychronization, giving rise to repeating spatiotemporal patterns, some of which form circuits that encode hierarchical feature relationships. Notably, these circuits are robust, even with the randomised, Poisson-distributed spike timings that represent the visual stimuli in the input layer. These results provide evidence for binding by polychrony as a feasible solution to the feature binding problem, and characterise the mechanism by which it may function. This mechanism can guide experimentalists in identifying such circuits in vivo, and could also be utilised in computer vision systems to capture more information and improve robustness to adversarial inputs. Introduction Overview and aims Eguchi, Isbister, and colleagues [1,2] have advanced a theory of hierarchical feature binding. These authors described how the feature binding problem, in the context of vision, concerns how the visual system represents a hierarchy of features (from edges to objects) at different spatial scales, and how it encodes the binding relationships between them (e.g., “this edge is part of that object”). Put another way, feature binding refers to how the visual system encodes which features are part of other features, dynamically linking them to form coherent object representations. Such hierarchical feature binding is thought to support the semantic interpretation of visual scenes, and may contribute to robust object recognition. Analogous hierarchical binding mechanisms are also likely to operate in other sensory domains such as audition. For example, in the human auditory system, phonemes may be bound to the words they form, which may in turn be bound to the identity of a particular speaker. Various theoretical frameworks address mechanisms of feature binding in the brain [3,4]. A remaining challenge is to specify how binding can be implemented in hierarchical neural circuits in a way that scales across spatial levels and remains compatible with realistic axonal delays and spiking dynamics. Although binding may also be supported within single neurons via nonlinear dendritic integration [5,6], we consider the complementary question of how “part-of” relations can be encoded at the level of population activity and between-layer interactions. The aim of this modelling study is to synthesise core aspects of prior related work. That is, we model the emergence of visual neurons representing a hierarchy of boundary contour elements of 2D shapes as previously performed by [7], but within more biologically realistic spiking neural network (SNN) models [1,2]. This enables us to investigate the simultaneous emergence of a hierarchy of boundary contour features, polychronization across network layers, and the formation of hierarchical feature binding (HFB) circuits containing binding neurons that relate contour features across increasing spatial scales. The use of a structured boundary-shape ensemble, which we describe as an ecologically realistic dataset in the sense of its feature statistics, provides a systematic test case for these mechanisms, and enables closer comparison with neuronal response properties reported in experimental studies [8–10]. Our primary advance beyond [1,2] is to move from a proof-of-principle demonstration of hierarchical binding by polychronization to a systematic test of the mechanism under a larger and more structured object ensemble. Specifically, we train and analyse the same type of hierarchical SNNs on diverse boundary-shape datasets, and we quantify how feature selectivity and polychronization co-develop across layers. We also focus explicitly on three-neuron HFB circuits, using their circuit-level timing constraints to identify binding relationships and to characterise where and how such motifs emerge throughout the hierarchy. We next provide biological motivation for hierarchical feature representations and feature binding in the primate ventral visual system, before outlining the modelling context and the specific binding-by-polychrony mechanism examined here. Hierarchy and feature binding in biological vision Human visual experience is hierarchical in nature [11]. That is, we perceive a scene as composed of a hierarchy of discrete but related visual features and objects at different spatial scales. For example, when we look at a face, we can identify constituent features such as the eyes and mouth as discrete features which are separate but related by the fact that they are part of the face, which is itself a larger scale discrete object. This cognitive process of binding smaller scale visual features to the larger scale visual features of which they are a part operates at every spatial scale across the visual field. Such semantically rich hierarchical feature binding underpins our ability to interpret and make sense of complex natural scenes. The term “hierarchy” is used in several distinct ways in systems neuroscience, and this can invite overinterpretation if left implicit [12]. In this paper, we use hierarchy primarily in a functional sense: representations become progressively more abstract across successive processing stages, with features at smaller spatial scales providing constituent parts for features at larger scales. In the taxonomy of [12], this aligns most closely with a progression of scales that describes nested feature representations. As such, our model targets how multi-scale feature representations can be learned and related by binding operations across stages, rather than making claims about any single anatomical definition of hierarchy. The need for explicit binding mechanisms is highlighted by classic theoretical arguments. When multiple objects are present, distributed feature representations can become ambiguous, entailing the “superposition catastrophe” in which the system cannot determine which features belong together [13,14]. Early accounts such as feature integration theory proposed that attention binds features at a single spatial locus [15], while binding-by-synchrony instead posits that neurons representing the same object fire in a temporally coordinated manner [16,17]. However, synchrony-based binding is challenged by the fact that biologically realistic heterogeneity in axonal delays tends to degrade precise synchrony, motivating alternative timing-based schemes. At the same time, neurophysiology provides direct support for relational and temporal codes. Border-ownership cells in the early visual cortex (e.g., area V2) respond to an oriented edge while also signalling which side belongs to a figure, effectively encoding a binding relation between a local contour and an object-level assignment [18]. More generally, stimulus-specific information is often present in reliable spatiotemporal firing patterns, suggesting that precise spike timing can serve as a substrate for binding beyond rate codes [19,20]. These observations motivate mechanistic accounts in which binding is expressed through reproducible spike-time relationships, such as polychronous patterns that can operate with heterogeneous conduction delays [21]. Ventral pathway feature hierarchy and boundary representations Here we consider the primate visual cortex, a series of interconnected cortical areas organised hierarchically to process visual information at increasing levels of abstraction. Within the primate ventral visual pathway, “hierarchy” is often described in terms of a systematic change in receptive-field scale and in the complexity of represented visual structure as signals progress from early to later areas. In early cortex (e.g., V1), neurons respond selectively to local oriented edges, while intermediate areas such as V4 contain neurons tuned to boundary curvature in an object-centred frame, and still later regions in inferotemporal cortex can integrate information over larger extents to support selectivity for more complex shape structure [8,10,11]. Visual signals in the primate cortex are often described as being processed through two distinct pathways: the dorsal stream, which rapidly processes motion and low spatial frequency signals, and the ventral stream, which processes colour and high spatial frequency signals at a slower speed. The latter is responsible for object recognition and representation of form [22] and is therefore the system of interest in this present study. The ventral visual pathway is depicted in Fig 1A with key cortical areas labelled. Cortical areas and boundary contour feature representations in the primate ventral visual pathway. (A) Cortical areas in the primate ventral visual pathway. As the pathway is ascended, neurons are selective for increasingly complex and larger scale shape features. (B) V4 neurons are selective to localised boundary contour elements in the frame of reference of an object. (C) Neurons in TEO and posterior TE encode larger scale combinations of localised boundary contour elements. Fig 1A adapted from a figure originally published in [23] under CC BY 4.0. In the initial stage of cortical processing within area V1, simple cells are selective for oriented bars and edges, complemented by complex cells that exhibit similar selectivity but additionally with a degree of translation invariance [24]. Following area V1, the visual signals proceed through areas V2, V4, TEO, and TE [25]. Ascending this pathway, the size of neuronal receptive fields increases, and neurons are selective for increasingly complex spatial features. In particular, lower cortical areas V1, V2, and V4 appear to process smaller scale visual features such as edges, contours, and colour, while the higher cortical areas TEO and TE are implicated in the processing of object shapes and faces. For example, in humans, the ventral temporal cortex includes the fusiform face area, which responds selectively to faces [26]. In particular relevance to this study are neurons found in area V4 that represent the boundary shapes of 2D objects, which encode the curvature of local boundary contour elements in an object-centred frame of reference [8,9]. Fig 1B illustrates the kind of localised boundary contour features that such a V4 neuron might respond to. Moreover, a population of such neurons can provide a distributed representation of the complete boundary shape of any 2D object. Ascending to the next areas of this visual pathway, namely TEO and posterior TE, neurons appear to integrate information from multiple V4 neurons in order to represent combinations of local boundary contour elements [10]. Fig 1C illustrates a combination of localised boundary contour features that a neuron in TEO or posterior TE might respond to. Finally, neurons at the end of the ventral visual pathway that encode whole objects may learn to respond to particular distributed representations of object boundary shape in earlier areas [27]. Related modelling work In a previously published modelling study [7], it was demonstrated how a rate-coded multi-layer neural network model of the ventral visual pathway could self-organise and develop neurons that encode the boundary contour elements of shapes in a manner similar to that observed in the visual brain [8–10]. That is, the neural network model developed a hierarchical representation of boundary contour elements at increasing spatial scales through successive layers of the network. To achieve this, the authors trained their model on a wide collection of distinct 2D object shapes. For most of these tasks, the network was trained on a set of 2D shapes formed from unique combinations of boundary contour elements: each shape consisted of n sides, and each side assumed one of p possible conformations. For example, a set of shapes described by n = 4 and p = 2 consisted of four-sided shapes, where each side would be either concave or convex. The image sets devised in [7] included all possible shape configurations, giving a total of pn images and p×n distinct boundary contour elements. Trained on such images, the synaptic weights within the network self-organised through a process of unsupervised competitive learning. Eguchi et al. [7] hypothesised that if the network model was trained on such a large number of different shapes, any particular boundary contour element, such as a right-side convexity, would only be seen with another particular boundary contour element, such as a bottom-side concavity, across a relatively small proportion of the full set of training images. This effectively creates a statistical decoupling between any two such boundary elements, which makes it less likely that a neuron in the network will learn to represent the combination of the two boundary contour elements. Moreover, as the total number of object shapes on which the network is trained increases, the total number of object shapes, pn, grows much larger than the total number of boundary contour elements, p×n, used to construct the objects. In this case, the capacity limit of a competitive neural network forces individual neurons to switch from representing whole object shapes to instead representing the boundary elements. In this situation, it was postulated that the unsupervised competitive learning mechanism operating within the network during training would force some neurons in the higher layers of the network to learn to represent particular boundary contour elements rather than the whole object shapes, which was indeed observed from the simulations. Moreover, the network model developed a hierarchy of neurons encoding boundary contour elements at increasing spatial scales through successive neuronal layers, as observed in the brain. Despite this reproduction of key aspects of ventral-stream shape processing, these simulations were carried out within a rate-coded neural network rather than an SNN. As such, the study did not incorporate temporally precise spiking dynamics, and it could not directly examine spike-timing-based binding mechanisms or the emergence of binding-neuron motifs that might encode hierarchical binding relationships between boundary contour elements across spatial scales. Hierarchical binding by polychronization Polychronization is a phenomenon observed in SNNs, characterised by groups of neurons firing in time-locked but asynchronous spatiotemporal patterns which recur [21]. These neuron assemblies, known as polychronous neuronal groups (PNGs), exhibit precise timing relationships despite the lack of simultaneous activation. This capability suggests a mechanism for encoding information, and offers a potential substrate for hierarchical feature binding relationships via coordinated activity in multi-layer SNNs. In addressing the hierarchical feature binding problem, the authors of [1,2] implemented a biologically realistic SNN model of the primate ventral visual pathway, characterised by several key properties: Their model explicitly simulated the precise timings of individual spikes fired by biologically-realistic neurons. Their model explicitly simulated the precise timings of individual spikes fired by biologically-realistic neurons. The synaptic weights between pairs of excitatory neurons were adapted during visually-guided learning in an unsupervised manner, dependent on the relative timing differences between emitted pre- and postsynaptic spikes: a phenomenon known as spike timing dependent plasticity (STDP). Long-term potentiation (LTP) of a synapse occurred in response to pre- before postsynaptic firing, and conversely long-term depression (LTD) for post- before presynaptic firing [28,29]. The synaptic weights between pairs of excitatory neurons were adapted during visually-guided learning in an unsupervised manner, dependent on the relative timing differences between emitted pre- and postsynaptic spikes: a phenomenon known as spike timing dependent plasticity (STDP). Long-term potentiation (LTP) of a synapse occurred in response to pre- before postsynaptic firing, and conversely long-term depression (LTD) for post- before presynaptic firing [28,29]. The network architecture was hierarchical, consisting of multiple, spatially-organised layers of neurons that included bottom-up, top-down and lateral synaptic connections. This kind of architecture is consistent with what is observed in the primate visual cortex. The network architecture was hierarchical, consisting of multiple, spatially-organised layers of neurons that included bottom-up, top-down and lateral synaptic connections. This kind of architecture is consistent with what is observed in the primate visual cortex. The network model incorporated axonal conduction delays in the synaptic connections between neurons, which governed the amount of time taken for a spike to propagate from its presynaptic origin to its postsynaptic target. The conduction delays between connected excitatory neurons were randomised at initialisation to values on the order of milliseconds to tens of milliseconds, consistent with measurements of biological networks [21]. The network model incorporated axonal conduction delays in the synaptic connections between neurons, which governed the amount of time taken for a spike to propagate from its presynaptic origin to its postsynaptic target. The conduction delays between connected excitatory neurons were randomised at initialisation to values on the order of milliseconds to tens of milliseconds, consistent with measurements of biological networks [21]. Driven by STDP, and presented with Poisson-encoded images, the SNN self-organised to produce higher-layer PNGs. These emergent patterns were interpreted as representing hierarchical binding relationships, linking visual features across multiple spatial scales. These prior studies were intentionally positioned as a proof-of-principle. Their stimulus sets consisted of three shapes, and the analysis emphasised demonstrating that polychronous structure and binding-neuron motifs can emerge in a biologically grounded hierarchical SNN [1,2]. In addition, much of the quantitative representational analysis in [1] focused on spike-pair PNGs. As noted in [2], apparent spike-pair structure can in some cases be difficult to disambiguate from elevated firing rates, motivating circuit-level criteria for identifying binding relationships. The authors of [1,2] hypothesised that training a spiking neural network with the above properties on a simple composite shape (e.g., the letter ‘T’) would lead to the formation of a PNG in the upper layers that responds selectively to that shape. Within such a PNG, they proposed that one or more binding neurons would emerge whose spiking indicates whether specific lower-level feature neurons contribute to activating higher-level feature neurons. Concretely, a binding neuron would fire as part of the PNG if and only if activity representing a constituent element (e.g., a vertical bar) is participating in driving activity representing the composite feature (e.g., the ‘T’), thereby signalling that this lower-level feature is part of the higher-level one. The simplest example of how such binding neurons might operate is shown in Fig 2. The hierarchical feature binding (HFB) circuit is comprised of three neurons: neuron 1 is located in a lower layer and represents a low-level feature (e.g., a vertical bar), neuron 2 is located in a higher layer and represents a high-level feature (e.g., the letter ‘T’), and neuron 3 is a binding neuron (which could be located in either of the two layers) that encodes the hierarchical binding relationship between these features. The connections between the three neurons have axonal conduction delays Δij, where j indexes the presynaptic neuron and i the postsynaptic neuron. Let us assume that the three axonal delays shown in Fig 2 satisfy where Δij is the time taken for a spike fired by neuron j to reach neuron i. In this case, the spikes fired by neurons 1 and 2 will arrive at the binding neuron 3 simultaneously if and only if neuron 1 (representing the low-level feature) is participating in driving neuron 2 (representing the high-level feature). If the neurons have fast synaptic time constants, then binding neuron 3 will fire if and only if the incoming spikes from neurons 1 and 2 arrive simultaneously. The result of all this is that the binding neuron 3 will fire if and only if the low-level feature encoded by neuron 1 is actually part of the high-level feature or object encoded by neuron 2. In this way, the binding neuron 3 comes to encode information about the HFB relationship between the lower-level and higher-level visual features. Simple theoretical binding circuit. This three-neuron hierarchical feature binding (HFB) circuit is the simplest kind of circuit posited to exist in the theory of hierarchical feature binding by polychronization [1,2]. Neuron 1 represents a low-level shape feature such as a vertical bar, neuron 2 represents a higher-level shape feature such as the letter ‘T’, and neuron 3 is a ‘binding neuron’ that represents whether the vertical bar is part of the ‘T’. Binding neuron 3 thus encodes the hierarchical binding relationship between the two features represented by neurons 1 and 2. The binding neuron 3 derives its firing characteristics from the pattern of axonal conduction delays between the three neurons. If Δ31≈Δ21+Δ32, then the spikes from neurons 1 and 2 will arrive at neuron 3 simultaneously and cause it to fire if and only if neuron 1 is participating in driving neuron 2. In this case, the activity of binding neuron 3 will encode that the lower-level feature represented by neuron 1 is part of the higher-level feature represented by neuron 2. It was found that three-neuron binding PNGs of the general form shown in Fig 2 developed naturally in the higher layers of their SNN model when the network was trained on images of objects using STDP to adapt the strengths of synaptic connections between excitatory neurons [1,2]. The present study addresses the main gaps highlighted by [1,2] without expanding the stimulus scope to translation invariance or multi-object scenes. We perform a systematic quantitative analysis of three-neuron HFB circuits as the unit of binding, and examine their emergence across all network layers. Crucially, we test these mechanisms under a larger and more structured boundary-shape ensemble, enabling us to measure how HFB-structured PNGs relate to the hierarchy of learned feature selectivity and to quantify their prevalence and layer-wise organisation. Study questions Specific questions we investigate through our simulations include the following: Does the network model develop a hierarchy of feature neurons, through unsupervised competitive learning, which respond selectively to boundary contour elements of increasing spatial scale as seen in neurophysiology studies? Does the network model develop a hierarchy of feature neurons, through unsupervised competitive learning, which respond selectively to boundary contour elements of increasing spatial scale as seen in neurophysiology studies? Do groups of neurons emerge in the higher network layers that exhibit the property of polychronization - i.e., emit their spikes in regularly repeating spatiotemporal patterns? Do groups of neurons emerge in the higher network layers that exhibit the property of polychronization - i.e., emit their spikes in regularly repeating spatiotemporal patterns? Do these neurons form HFB circuits, of the form depicted in Fig 2, which encode the hierarchical binding relationships between boundary contour elements at different spatial scales across the network hierarchy? Do these neurons form HFB circuits, of the form depicted in Fig 2, which encode the hierarchical binding relationships between boundary contour elements at different spatial scales across the network hierarchy? Materials and methods Hierarchical network architecture The hierarchical SNN model explored in this study is depicted in Fig 3, and is designed to simulate the successive stages of visual information processing in the primate ventral visual pathway. This architecture derives from previous SNN modelling studies of this pathway [1,2], including notable studies of VisNet [7,30]. Specifically, it models four ascending cortical layers: V2, V4, TEO and TE, each critical to the visual processing pathway. Architecture of the SNN model of the primate ventral visual pathway. Processing begins with input images being convolved with a set of Gabor filters that mimic the responses of bar and edge detecting simple cells in cortical area V1. The outputs of these filters are then used to determine the firing rates of retinal Poisson neurons in layer L0, where the individual spike timings are randomly generated according to a Poisson distribution. This input layer feeds into a hierarchical structure comprising four layers of spiking neurons (layers L1 to L4), each containing distinct populations of excitatory and inhibitory neurons. There are bottom-up (feedforward) and top-down (feedback) excitatory synaptic connections between adjacent layers, lateral (recurrent) connections within layers among excitatory neurons, and bidirectional connections between excitatory and inhibitory neurons within the same layer. The feedforward and feedback connections between excitatory neurons in adjacent layers are topologically organised. That is, excitatory neurons receive connections only from topologically corresponding local regions of the adjacent layers. Lateral synaptic connections between excitatory neurons and between excitatory and inhibitory neurons within the same layer are also formed in localised regions. This topological organisation facilitates a scalable receptive field across the hierarchy of layers, with neurons in higher layers integrating inputs over increasingly larger spatial extents, depicted as neurons shaded in dark grey. Each of these layers consists of subpopulations of excitatory and inhibitory interneurons, interconnected through certain synaptic projections. In more detail, connections between excitatory neurons are modifiable, and between adjacent layers there exist bottom-up (feedforward) and top-down (feedback) connections. Within the same layer, lateral connections exist between excitatory neurons. These connections are topologically organised, such that excitatory neurons only receive connections from neurons in a locally corresponding region of the same or an adjacent layer, which is determined according to the extent of their receptive fields. In this context, the receptive field of a particular neuron is a localised circular region of the visual space that affects its firing. This circular region is defined by a fan-in radius applied for each layer between the neuron and the visual space, along with an associated connection probability. This model implements lateral competition among excitatory neurons, which is mediated by the subpopulation of inhibitory interneurons. This competition encourages feature selectivity and information sparsification, mimicking the competitive dynamics observed in the primate cortex, with a biologically accurate ratio of excitatory to inhibitory neurons close to 4:1. This lateral competition is supported via bi-directional synaptic connections between excitatory and inhibitory neurons, which are non-modifiable and formed between neurons in close proximity with each other. Taken together, these topological connections are designed to promote increasingly complex and abstract representations of visual stimuli as information ascends through the visual pathway: the fan-in radius of connections from the retina expands in successive layers, allowing for a broader integration of visual features at higher levels of processing. With respect to all network layers the feedforward, feedback and lateral connections between excitatory neurons incorporate axonal conduction delays with values on the time-scale of milliseconds, within biologically realistic constraints. The conduction delay of a synaptic connection refers to the time taken for an emitted spike to propagate from a presynaptic neuron and reach its postsynaptic target. Excitatory-excitatory connections are assigned conduction delays with random values selected from a uniform distribution, ranging from 1–10 ms, whereas connections associated with inhibitory neurons are fast and assigned a conduction delay of 0.1 ms. Introducing randomised conduction delays between excitatory neurons is essential to support the emergence of polychronization, and by extension feature binding relationships in the network. This is because these randomised delays provide a sufficient variety of initial neuronal structures for presented visual stimuli to activate and subsequently self-organise as some circuits are reinforced through STDP. For detailed numerical parameters, including the numbers of excitatory and inhibitory neurons, the specific fan-in radii, associated connection statistics with respect to each layer, axonal conduction delays, and other topological characteristics of the synaptic connections, refer to Table 1. 10.1371/journal.pcbi.1014752.t001Table 1Model parameter choices. Terms are abbreviated as follows: FF = feedforward, FB = feedback, LAT = lateral, STDP = spike-timing-dependent plasticity. Multiple parameter choices inside curly braces are selected according to the run experiment, and specified in the results.ParametersSymbolLayer1234 (a) Network parameters Number of excit. neurons– 64×64 64×64 64×64 64×64 Number of inhib. neurons– 32×32 32×32 32×32 32×32 Average number of FF afferent excit. connections per excit. neuron–50100100100Fan-in radius for FF afferent excit. connections to each excit. neuron–281216Average number of FB afferent excit. connections per excit. neuron–{0, 10}{0, 10}{0, 10}–Fan-in radius for FB afferent excit. connections to each excit. neuron–888–Average number of LAT afferent excit. connections per excit. neuron–{0, 10}{0, 10}{0, 10}{0, 10}Fan-in radius for LAT afferent excit. connections to each excit. neuron–2222Average number of LAT afferent excit. connections per inhib. neuron–10101010Fan-in radius for LAT afferent excit. connections to each inhib. neuron–2222Average number of LAT afferent inhib. connections per excit. neuron–30303030Fan-in radius for LAT afferent inhib. connections to each excit. neuron–4444 (b) Gabor filter parameters Phase shift ϕ −π/2,π/2 Wavelength λ 0.8Orientation θ 0,π/4,π/2,3π/4 Spatial bandwidth b 1Aspect ratio γ 0.5Kernel size (pixels) k 6 (c) Cellular parameters (excit. and inhib., respectively) Membrane capacitance Cm 500 pF, 214 pFLeakage conductance gL 25 nS, 18 nSMembrane time constant τm 20 ms, 12 msResting potential V 0 -74 mV, -82 mVReversal potential V^ 0 mV, -70 mVAdaptation reversal potential for excit. cells V^a -90 mVFiring threshold potential Θ -53 mV, -53 mVAfter-spike hyperpolarization potential Vh -57 mV, -58 mVAbsolute refractory period τR 2 msPoisson neuron firing rate scaling factor β 6 ms–1 (d) Synaptic parameters Pre- and postsynaptic activity trace scaling factors αC,αD 0.5Pre- and postsynaptic STDP time constants τC,τD 5 msLearning rate ρ {0.02, 0.04}Maximum excit. and inhib. axonal conduction delays–10 ms, 0.1 msConductance scaling factor for FF connections λEfE 30 nSConductance scaling factor for FB connections λEbE 20 nSConductance scaling factor for LAT excit. to excit. connections λElE 20 nSConductance scaling factor for LAT excit. to inhib. connections λElI 30 nSConductance scaling factor for LAT inhib. to excit. connections λIlE 20 nSConductance scaling factor for adaptation in exc. cells λa 6 nSSynaptic time constant for all excit. to excit. connections τg{EfE,EbE,ElE} 2 msSynaptic time constant for LAT excit. to inhib. connections τgElI 2 msSynaptic time constant for LAT inhib. to excit. connections τgIlE 5 msSynaptic time constant for adaptation in exc. cells τa 80 ms (e) Simulation parameters Exposure time per image (training)–100 msExposure time per image (inference)–250 msNumerical time step Δt 0.1 ms Model parameter choices. Terms are abbreviated as follows: FF = feedforward, FB = feedback, LAT = lateral, STDP = spike-timing-dependent plasticity. Multiple parameter choices inside curly braces are selected according to the run experiment, and specified in the results. Preprocessing visual stimuli by Gabor filtering Prior to the network processing visual stimuli, images are first transformed using Gabor filters to simulate the tuning profiles of simple cells in the primary visual cortex (V1). The filter parameters are selected to align with the firing properties of V1 cells, aimed at extracting localised orientations of shape boundaries within the image space [31,32]. This approach ensures a distinct pattern of filter outputs for each of the visual objects, providing a suitable basis for generating input spikes by Poisson neurons in the network’s initial, retinal-V1 layer. Specifically, the filters are defined by the following set of equations: where x and y specify the position of a light impulse in the visual field [33]. The other parameters include λ as the wavelength, θ the feature orientation, ϕ the phase, b the spatial bandwidth in octaves and γ the aspect ratio. The values of λ,θ,ϕ,b,γ as well as the kernel size k used to construct the Gabor filters are given in Table 1. The total number of Gabor filters implemented within the SNN architecture is determined by the number of distinct parameter combinations used to construct the filters. Additionally, to balance the filter outputs and to assign them equal importance, these filters have their mean set to zero and their Euclidean norm set to one. Using these constructed Gabor filters, a grayscale image is preprocessed and converted into its filtered form as follows [34]: The image is resized to a target resolution of 128×128 pixels, matched to the spatial dimensions of the network’s initial layer. The image is also standardised by dividing it by its standard deviation. The image is resized to a target resolution of 128×128 pixels, matched to the spatial dimensions of the network’s initial layer. The image is also standardised by dividing it by its standard deviation. These Gabor filters are independently convolved with the image in two dimensions, maintaining the original image size. This convolution produces a retinotopic 2D map reflecting filter output values across the image, where the different filter types are vertically stacked as separate channels. These Gabor filters are independently convolved with the image in two dimensions, maintaining the original image size. This convolution produces a retinotopic 2D map reflecting filter output values across the image, where the different filter types are vertically stacked as separate channels. Each channel is rescaled by dividing its filter outputs with their Euclidean norm, thus ensuring a balanced input representation. Each channel is rescaled by dividing its filter outputs with their Euclidean norm, thus ensuring a balanced input representation. Following the above steps, these filter outputs are used to determine the firing rates of Poisson neurons constituting the network’s initial layer L0, with a direct one-to-one correspondence between each neuron and filter output. Hence, the input firing rates have a dimensionality of ℝW×H×C, where W and H refer to the spatial dimensions of the input space, being 128×128, and C the number of filter channels. Specifically, Poisson spike trains are generated with the probability of a spike event occurring over a small interval [t,t+Δt] given by: where f indexes the filter channel, β is a scaling factor which is tuned to constrain the input firing rate to less than 100 Hz, and Δt≪1 ms is a small time step. These Poisson neurons have an excitatory action, and project to the subpopulation of excitatory neurons residing in layer L1 in a topologically corresponding manner (Fig 3): each excitatory neuron in L1 receives its input from a spatially aligned, localised region of Poisson neurons, spanning all of the filter channels, with a fan-in radius and expected number of connections given by Table 1. We hypothesise that training SNNs on randomised (Poisson) input spike timings will cause the synaptic connections in the higher layers to self-organise via STDP to produce visual representations that are highly robust with respect to image noise. The randomness introduced at the input layer will force the network to learn patterns without relying on the precise timing of individual spikes. This could lead to the emergence of synaptic connections that are more invariant to noise and perturbations. Differential equations Leaky-integrate-and-fire model. Following related SNN modelling studies [1,2], neurons in all layers of our network except V1 are modelled as conductance-based leaky-integrate-and-fire (LIF) neurons [35]. According to this model, the membrane potential Vi of a neuron i at time t is updated as follows: where Vi is driven by synaptic current Ii integrated over the neuron’s afferent synapses; the membrane potential is driven upwards in response to current from excitatory synaptic conductances and downwards with respect to inhibitory synaptic conductances. The membrane properties of the neuron determine its decay back towards its resting state, described by its resting potential V0. The membrane time constant is defined by τm=Cm/g0, where Cm and g0 refer to the membrane capacitance and leakage conductance, respectively. The neuron’s membrane resistance is the inverse of the leakage conductance, given by R = 1 / g0. The neuron belongs to one of two classes: excitatory or inhibitory, with class-specific parameter values indexed by the superscript γ. Concerning spike generation, the neuron fires a spike at time tif when its membrane potential crosses the firing threshold Θ from below. Immediately after the neuron fires, its membrane potential is reset to its after-spike hyperpolarisation potential Vh where it is held for an absolute refractory period τR. The total synaptic current entering a neuron is calculated as the sum of all afferent synaptic conductances, both excitatory and inhibitory, multiplied by the potential difference between the reversal potential specific to the synapse class, V^γ, and the membrane potential of the neuron, Vi. The time-dependent conductance for a particular synapse is denoted by gij with j and i indexing the pre- and postsynaptic neurons, respectively: Additionally, for an excitatory postsynaptic neuron i, the total synaptic current is adjusted to include an adaptation conductance, gia, with an associated reversal potential V^a as follows: This homeostatic mechanism is necessary to regulate the firing activity of excitatory neurons, preventing excessively high firing rates from emerging during network training [36]. Synaptic conductance equations. The conductance of a specific synapse, gij, is determined by two components: a decay term τg, and a change induced by a Dirac delta function when spikes arrive from a presynaptic neuron j. This relationship is expressed as follows: where the axonal conduction delay associated with a particular synapse is denoted by Δtij, and the received presynaptic spikes are indexed by l. The conduction delays are initialised by either selecting random, uniformly distributed values ranging between 1–10 ms for excitatory synapses, or being set to 0.1 ms for inhibitory synapses. The parameter λ is a biological scaling constant, and is used to scale the synaptic efficacy (or weight) Δgij which lies between unity and zero. The Dirac delta function captures the instantaneous change in gij at the time of presynaptic spike arrival, and is defined as follows: Similarly, the dynamics of the adaptation conductance, gia, for an excitatory postsynaptic neuron i is determined as follows: where the parameter λa determines the size of the instantaneous change in gia at the time of an emitted postsynaptic spike, which is indexed by k. Synaptic learning equations. The differential equations outlined below characterise STDP at modifiable excitatory-to-excitatory synapses, where plasticity is implemented across all bottom-up, top-down, and lateral connections among excitatory neurons spanning layers 1–4 [1,2]. These STDP equations are adapted from [37], and are described for a synaptic connection between pre- and postsynaptic neurons j and i, respectively, as follows. The recent presynaptic activity is tracked by a variable, Cij, and can be understood as the concentration of neurotransmitter (glutamate) released into the synaptic cleft [37]. This activity variable is constrained within the closed interval [0, 1] for 0≤αC≤1, and is governed by a decay term τC as follows: This activity trace is driven up in response to the arrival of presynaptic spikes at the postsynaptic neuron, scaled by αC, where the conduction delay associated with the synapse is accounted for via Δtij; hence, Cij is only updated when presynaptic spikes actually arrive at the postsynaptic neuron, as opposed to when they are originally fired by the presynaptic neuron. The postsynaptic neuron’s recent activity, tracked by the variable Di, reflects the proportion of unblocked NMDA receptors due to recent depolarization from back-propagated action potentials [37]. The time evolution of this activity variable is controlled by a decay term τD, and is modelled as follows: where Di is driven up in response to the emission of postsynaptic spikes by an amount proportional to the parameter αD. Postsynaptic spike times are indexed by k. The strength of the synaptic weight, Δgij, is modified according to captured values for Cij and Di as follows: where τΔg is the time constant, and the learning rate, controlling the magnitude of a synaptic weight change, is denoted by ρ. The summation index k is over spikes emitted by the postsynaptic neuron i, and the summation index l is over spikes emitted by the presynaptic neuron j. For the rest of this paper, the learning rate is redefined as ρ←ρ/τΔg, thus eliminating τΔg from the above equation to more directly reflect the effective weight change arising from STDP. The described STDP model functions in the following manner. The first term on the RHS of Eq (12) represents long-term potentiation (LTP): if the presynaptic activity variable Cij is elevated, indicative of recent presynaptic spikes at the postsynaptic neuron, and a postsynaptic spike occurs, synaptic potentiation is facilitated. This results in an increase in the synaptic weight. Conversely, the second term on the RHS of Eq (12) represents long-term depression (LTD): if the postsynaptic activity variable Di is elevated, reflecting recent postsynaptic spiking, and a presynaptic spike arrives at the postsynaptic neuron, synaptic depression occurs. This leads to a reduction in the synaptic weight. The parameter values used in this paper are specified in Table 1, and parameters with multiple choices, such as the learning rate ρ, are described in the results. Numerical scheme. The differential equations previously outlined are discretised into finite difference equations and solved using a fourth order Runge-Kutta (RK4) integration method. Simulations are conducted with a time step of Δt=0.1 ms, using the Brian2 simulator [38]. Topological connectivity Every neuron in the network is assigned a particular spatial location within its layer. The Poisson neurons in L0 are each assigned to a specific retinotopic location within the visual field. Each Poisson neuron then represents a particular form of local edge or bar feature at the corresponding retinotopic location within an image. Neurons in the higher layers 1–4 are also arranged in a corresponding retinotopic manner within their respective layers. Individual excitatory neurons within layers 1–4 receive bottom-up and top-down connections from excitatory neurons within topologically corresponding localised regions of the adjacent layers as illustrated in Fig 3. Excitatory neurons in adjacent layers are connected with probability pconn (set per projection class to match the expected number of afferent connections in Table 1), if the condition specified by Eq (13) is met: where x→i and x→j are respectively the spatial locations of the post- and presynaptic neurons, x is horizontal location, y is vertical location, and rin is the fan-in radius. Implementing bottom-up connectivity with this kind of topological organisation allows the development of classical neuronal receptive fields that increase in size as the network layers are ascended. Synaptic connections within layers are formed using the same distance criterion (Eq (13)): each excitatory neuron receives lateral excitatory input from nearby excitatory neurons (E→E), and lateral inhibitory input from nearby inhibitory interneurons (I→E), while inhibitory interneurons receive local drive from nearby excitatory neurons (E→I). These lateral projections are sampled with the corresponding pconn within their fan-in radii, with expected connection counts per neuron specified in Table 1. Unsupervised competitive learning The SNN model presented in this paper is trained using unsupervised competitive learning. In contrast to the kind of supervised learning that is typically used to train deep neural networks for machine vision tasks, the network is not supplied with any additional information about the presented visual stimuli, such as object shape labels, during training. In particular, we do not instruct the neurons in the final layer how they should respond to each of the shape images during training. Unsupervised competitive learning operates as follows: During training, each image is presented in turn to the network. During each image presentation, activity is allowed to flow through the synaptic connectivity within the network for several hundred milliseconds. During training, each image is presented in turn to the network. During each image presentation, activity is allowed to flow through the synaptic connectivity within the network for several hundred milliseconds. Within each layer there are separate pools of excitatory and inhibitory neurons. There are bi-directional connections between these excitatory and inhibitory pools. The excitatory neurons excite the inhibitory neurons, while the inhibitory neurons depress the activity of the excitatory neurons. Within each layer there are separate pools of excitatory and inhibitory neurons. There are bi-directional connections between these excitatory and inhibitory pools. The excitatory neurons excite the inhibitory neurons, while the inhibitory neurons depress the activity of the excitatory neurons. When a subpopulation of excitatory neurons within a layer becomes activated due to the visual input from an image presented to the network, they excite a subpopulation of inhibitory neurons in the same layer. These active inhibitory neurons then depress the activity of nearby excitatory neurons. In this way, the excitatory neurons effectively ‘compete’ with each other through lateral inhibition mediated by the pool of inhibitory neurons. The excitatory neurons that ‘win’ the competition and remain active then learn via STDP in the bottom-up connections to respond to the current active inputs from the preceding layer. This process occurs in every layer of the network, from the final layer to the input layer. When a subpopulation of excitatory neurons within a layer becomes activated due to the visual input from an image presented to the network, they excite a subpopulation of inhibitory neurons in the same layer. These active inhibitory neurons then depress the activity of nearby excitatory neurons. In this way, the excitatory neurons effectively ‘compete’ with each other through lateral inhibition mediated by the pool of inhibitory neurons. The excitatory neurons that ‘win’ the competition and remain active then learn via STDP in the bottom-up connections to respond to the current active inputs from the preceding layer. This process occurs in every layer of the network, from the final layer to the input layer. In this way, excitatory neurons throughout the network learn to respond to specific visual features present within the images at different spatial scales. Neurons in the lower layers, which receive signals from smaller regions of retinal space and so have smaller receptive fields, will typically learn to respond to smaller, simpler visual features such as short local boundary contour elements. While neurons in higher layers, with larger receptive fields, will typically learn to respond to larger, more complex features such as longer boundary contour elements. In this way, excitatory neurons throughout the network learn to respond to specific visual features present within the images at different spatial scales. Neurons in the lower layers, which receive signals from smaller regions of retinal space and so have smaller receptive fields, will typically learn to respond to smaller, simpler visual features such as short local boundary contour elements. While neurons in higher layers, with larger receptive fields, will typically learn to respond to larger, more complex features such as longer boundary contour elements. We demonstrate here that unsupervised competitive learning combined with STDP also leads to the emergence of ‘binding neurons’ that represent the HFB relationships between visual features at different spatial scales across successive layers. As stated, these outcomes occur without the need to specify how the neurons in the final layer should respond to each of the object shape images during training. Unsupervised competitive learning within SNNs may produce quite different kinds of visual features than those that emerge within deep neural networks trained by supervised learning. Unsupervised competitive learning drives neurons within the network to learn to represent the statistically significant visual features that tend to reoccur regularly across images. We propose that such natural visual features may help to provide a basis set for the representation of objects that is more robust with respect to image noise and natural object transformations. In contrast, the features that emerge within deep neural networks may be more abstract and lead to arbitrary and unnatural decision boundaries. Training on visual stimuli and inference During training the network is exposed to an ecologically realistic set of similar 2D object shapes, where each shape is configured from a unique combination of boundary contour elements [7]. Specifically, a shape is described by n sides and each side assumes one of p possible conformations, giving rise to a total of pn different shapes and p×n unique boundary contour elements. This study only considers shapes with two conformations at each side (convex or concave), and sets of images containing either three- or four-sided shapes. Hence, the network is trained on a set of images containing either 8 or 16 shapes, referred to by the dataset names ‘N3P2’ and ‘N4P2’, respectively (Fig 4). Sets of visual stimuli used for training, consisting of simple 2D object shapes formed from n sides and p conformations per side. (A) This reduced set, termed N3P2, consisted of shapes described by n = 3 and p = 2: totalling pn=8 different shapes and n×p=6 distinct boundary contour elements. (B) This expanded set, N4P2, was described by n = 4 and p = 2: totalling 16 shapes and 8 boundary contour elements. Because N3P2 and N4P2 enumerate all pn configurations with p = 2 (convex vs. concave per side), both datasets are exactly balanced. For each object side, convex and concave boundary elements occur equally often (50% each), and for any pair of sides all four joint configurations are equally represented. As a compact summary of feature diversity, the number of convex sides per shape follows the binomial counts: for N3P2 the counts for k = 0,1,2,3 convex sides are 1,3,3,1 (out of 8 shapes), and for N4P2 the counts for k = 0,1,2,3,4 are 1,4,6,4,1 (out of 16 shapes). These summaries clarify in what sense the datasets are “ecologically realistic” for our modelling objective: the same small set of boundary parts is reused across many distinct shapes, so each part is seen across multiple contexts. This is the condition required for the “statistical decoupling” argument, and it promotes the emergence of feature-based representations under competitive learning. The network’s learning rate, ρ, is set to 0.04 and 0.02 when trained on N3P2 and N4P2, respectively. The set of images for a dataset are presented to the network as separate epochs, where one training epoch corresponds to the presentation of each one of these images. The duration of an image presentation is 100 ms, and the total number of training epochs is 60, which was established from preliminary simulation runs to result in a convergent network state; this is equivalent to 48 s and 96 s of total simulation time with respect to the N3P2 and N4P2 datasets, respectively. The modifiable connections in the network which are adapted via STDP during training include feedforward (FF), feedback (FB) and excitatory-to-excitatory lateral (LAT) connections. Non-modifiable connections which remain fixed during training include the connections between excitatory and inhibitory neurons. The synaptic weights Δgij of modifiable connections are initialised by independently selecting values from a uniform distribution in the range [0, 1], while non-modifiable connections are initialised with synaptic weight values that are all set to unity. Concerning inference, after training STDP is disabled and the spiking activity of neurons across all layers is recorded in response to the same training images being presented to the network. Each image is presented to the network for a duration of 250 ms, and this is repeated a certain number of times to draw statistically meaningful results. Additionally, and unless specified otherwise, the first 50 ms of each recording is discarded to account for the spillover in network responses between successive image presentations, thereby allowing it time to settle into each new activity state. Performance measures Single cell information analysis. Information theory provides a framework to quantify the amount of information that individual neurons convey regarding specific types of visual stimuli through their firing rate responses, which can be used as a measure to assess how selectively neurons respond to their preferred stimuli. In this context, a specific type of stimulus refers to a set of object shapes sharing a common characteristic, such as a left-side convexity versus a left-side concavity. For example, all shapes with a left-side convexity can be considered as variations or “transforms” with respect to this characteristic, despite differences in the other visual features. Accordingly, the set of shapes sharing this feature represents the distinct forms, or transforms, of a specific visual stimulus, denoted as s [7]. To evaluate the amount of information each neuron carries about these distinct stimulus types, we compute an information-theoretic measure based on the firing rate responses recorded during network inference. From analysing the distribution of these responses with respect to each type of stimulus, the stimulus-specific information conveyed by a neuron can be calculated using the following equation [39]: where s is a specific stimulus, r is the neuronal response to s, and R→ is the set of neuronal responses across the entire set of stimuli. Here, P(r|s) is estimated empirically from the distribution of firing-rate responses over repeated presentations of shapes belonging to class s, and P(r) is the corresponding marginal across all stimulus classes. The maximum amount of information that a neuron can impart about n distinct types of stimuli is given by log2(n) bits, signifying an unambiguous response profile. This study considers two categories of visual stimuli at a given object side: convex vs. concave conformations. We therefore compute stimulus-specific information separately for each side, for which the maximum is one bit. These calculations are performed for each excitatory neuron in the network’s final layer (L4) and for each object side. To summarise selectivity in a manner that is independent of side location, we additionally define each neuron’s preferred side as the one yielding its maximum stimulus-specific information. When reporting convex-selective responses, this maximum value is retained only if the neuron’s mean firing rate is higher for convex than concave at its preferred side (otherwise the convex-selective information is set to zero); an analogous criterion is used when reporting concave selectivity by swapping the stimulus labels. This procedure allows us to identify and rank neurons that are most informative about the presented stimuli, either at a fixed side or using a pooled (max-over-sides) summary, depending on the analysis. Traceback analysis. Following the identification of neurons that are particularly informative and selectively responsive to distinct types of visual stimuli, a technique termed ‘traceback analysis’ is applied to locate the regions within the retinal input space (layer L0) to which these neurons are most sensitive. By examining these regions through visual inspection, the specific object shape features that are primarily responsible for eliciting the neuronal responses can be identified. This analysis is used to confirm that trained neurons become tuned to specific boundary contour elements and provides a mechanistic explanation for their observed firing rate characteristics. The procedure for this method is as follows. We start by considering an excitatory neuron indexed by i that resides in the final layer l = L, where L > 0. Hence, the neuron’s activation is defined as where σ is an activation function, and netiL is the neuron’s weighted input. For tractability, we approximate a neuron’s activation as arising solely from feedforward excitatory connections, such that where the summation runs over all presynaptic excitatory neurons indexed by j in layer L−1 that project to the postsynaptic neuron i, and we assume wijL:=ΔgijL. Accordingly, using Eqs (15) and (16), the sensitivity of the postsynaptic neuron’s activation to that of a selected Poisson neuron indexed k in the retinal layer (l = 0) can be expressed as follows: To compute the above expression, we apply the technique of backpropagation, defining sensitivity terms with respect to the last and preceding layers as follows: and recursively for intermediate layers, 1