Abstract
Cortical areas comprise multiple types of inhibitory interneurons with stereotypical connectivity motifs, but the combined effect of different inhibitory connectivity patterns on postsynaptic dynamics has been largely unexplored. Here, we analyse the response of a single postsynaptic neuron receiving tuned excitatory connections that are balanced by various combinations of inhibitory input profiles. Inhibitory tuning can be flat, share the same tuning preference as the excitation or, alternatively, it can feature counter-tuning such that non-preferred excitatory inputs receive large inhibition. When all inhibitory populations are active, the net inhibitory effect is the same regardless of the tuning profile. By modulating the activity of specific inhibitory populations, strongly correlated responses to preferred or non-preferred inputs, as well as uncorrelated responses emerge. Moreover, biologically inspired inhibitory plasticity rules produce the necessary connectivity profiles, indicating how plasticity rules in various cell types can interact to shape cortical circuit motifs and their dynamics.
Inhibitory neurons exhibit large variability in morphology, connectivity motifs, and electrophysiological properties1–4. One of inhibition’s main function is to balance excitatory inputs, thus stabilising neuronal network activity5,6, and allowing for a range of different functions to be implemented by the brain7–11. When both inhibitory and excitatory inputs share the same statistics, and their weight profiles (receptive fields) are similar12, the resulting state of the postsynaptic neuron is one of precise balance of input currents9. Modulation of inhibition, e.g., decrease or increase in local inhibitory activity and, consequently, change in the balance between excitation and inhibition, is essential to control the activity of neuronal groups13. It is thus believed that disinhibition is an important mechanism for the implementation of high-level brain functions, such as attention14,15, memory retrieval6,16,17, signal gating18,19, and rapid learning11. This type of modulation cannot be easily implemented, and may hint at why neuronal circuit motifs have so many interneuron types1,2,7.
To reach a state of precise balance, a Hebbian-like inhibitory plasticity rule — increase in synaptic weights for correlated pre- and postsynaptic activity, as observed, e.g., in auditory cortex20 — is required6. A form of anti-Hebbian inhibitory plasticity — decrease of synaptic weights for correlated pre- and postsynaptic activity — has also been reported21–23, and such a rule has been proposed as a mechanism for memory formation and retrieval24. These two types of plasticity rules would form opposite receptive fields for inhibitory connections: synapses following a Hebbian-like inhibitory plasticity rule would mirror excitatory inputs while an anti-Hebbian plasticity rule would impose strong inhibitory inputs for weak excitatory ones, and vice-versa. In line with this hypothesis, intracellular recordings indicate that strong inhibitory postsynaptic potentials can be elicited by stimuli with preferred orientations25,26, but also by stimuli with non-preferred orientations27,28. Additionally, postsynaptic responses in auditory cortex of mice have been shown to vary with sound intensity29.
To understand the mechanistic origin of such varying responses from the same cells, we investigated the behaviour of a single postsynaptic neuron model receiving tuned excitatory inputs, and inhibition from two distinctly tuned populations. Tuning may correspond to preference to a specific sound frequency12, orientation of visual cues30, taste31, whisker stimulation32, or position in space33. We show that when the postsynaptic neuron is in a balanced state with respect to its excitatory and inhibitory inputs, preferred signals are transiently revealed, but steady state responses are indiscriminate of the stimulus preference6 (i.e., its ‘orientation’, etc.), regardless of the inhibitory connectivity. We could substantially alter the response profile of the postsynaptic neuron by modulating the activity of either of the two presynaptic inhibitory populations, allowing for the propagation of facets of the input patterns that were previously quenched by inhibition. We thus introduce a mechanism to selectively filter stimuli according to, e.g., attentional cues. We show that the tuning profiles necessary for such attentional gating can be achieved by a set of biologically plausible plasticity rules. These rules are based on previous theoretical and experimental work, suggesting that distinct plasticity rules can harmoniously coexist in the brain. Our work proposes a simple biological implementation for an attentional switch at a mechanistic level, and provides a solution for how such a neuronal circuit can emerge with autonomous and unsupervised, biologically plausible plasticity rules.
Results
To study the effect of interacting populations of feedforward inhibition, we investigated the response of a single postsynaptic leaky integrate-and-fire neuron receiving tuned excitatory and inhibitory inputs. Excitatory inputs were organised into a single population, subdivided into 16 signal groups of 200 excitatory afferents. Inhibitory inputs initially formed a single population, mirroring the excitatory subdivision, but with 50 afferents per group. Later, we split the inhibitory inputs into two populations with 25 afferents per signal group (Fig. 1A, see Methods for details), allowing us to obtain two differently tuned populations (presumably types) of inhibition. Excitatory and inhibitory afferents belonging to the same group shared temporal fluctuations in firing rates, termed input patterns, even if they belonged to different populations. In our simulations, input patterns could either be natural or pulse. Natural inputs were generated through an inhomogeneous Poisson process based on a modified Ornstein-Uhlenbeck process (Fig. 1B,C), such that neurons of the same signal group also had temporally-correlated firing patterns (Fig. 1C, top; see also Ujfalussy et al.34 ). The resulting long-tail distribution of inter-spike-intervals (Fig. 1C, bottom) was similar to experimentally observed spike patterns in vivo35,6. We used this type of input to quantify steady-state (average) postsynaptic responses, and to train inhibitory synapses via plasticity rules.
In the alternative pulse input regime we analysed transient responses with 100-ms long pulses of varying amplitudes6. Pulses were delivered through a single signal group of excitatory and inhibitory afferents, while all other groups remained at baseline firing-rate (Methods). Responses were quantified according to postsynaptic firing rates during the first (phasic) and last (tonic) 50 ms stimulation (Fig. 1D), averaged over 100 trials. Separating responses in phasic and tonic allowed us to discriminate changes in output due to the input onset, and slower integration of the pulse, respectively.
Modulating a single inhibitory population
At first, we constructed a standard cortical circuit motif with one excitatory and one inhibitory population6,37 (Fig. 2A, top). Both excitatory and inhibitory weights were tuned according to a receptive-field-like profile, providing precise balance between excitation and inhibition (Fig. 2A, bottom) and average post-synaptic firing rates of 5 Hz for natural inputs (Fig. 2B, left; Fig. 2C, control). We then changed the gain of all inhibitory afferents by modulating their firing rates, from 50% to 150% of control rates. This change of input balance translated into changes in output rates (Fig. 2C, bottom), and spike patterns (Fig. 2B, middle and right). When inhibition was equal or larger than excitation, the output was largely uncorrelated to any given input signal (Fig. 2D, top). When inhibitory firing rates fell below 90% of the control condition, the output first began to correlate with the preferred input signal. When inhibition became even weaker, the correlations increased, and even non-preferred signals were articulated in the postsynaptic firing patterns (Fig. 2D, bottom).
Transient presynaptic activity pulses caused strong phasic responses in the balance state when they were delivered through the afferents of the preferred inputs (Fig. 2E, top row). Stimuli from non-preferred afferents were largely ignored. This discriminability between transients of low or high amplitude pulses decreased when inhibition was down-regulated (Fig. 2E, middle row) such that pulse stimuli from all signal groups caused a response. Increased inhibition, on the other hand, completely abolished transient responses to non-preferred afferents (Fig. 2E, bottom row). In all three cases (balanced control, weak and strong inhibition), the postsynaptic neuron elicited most of its spikes within the phasic period of the total 100 ms input step (Fig. 2E).
Co-tuned and flat inhibitory populations
Next we incorporated two types of inhibition into the circuit motif (Fig. 3A, top). One population mirrored the synaptic weight profile of the excitation. We tuned the inhibitory weights of this co-tuned population so that the output weights of non-preferred signal populations (far from the peak of the receptive field) decreased to zero. The other inhibitory population featured an un-tuned response profile (we termed this the flat population; Fig. 3A, bottom). We ensured that the sum of the inhibitory weights for each signal group balanced the excitatory inputs by default (Methods). We tested three conditions: both inhibitory populations active (control); co-tuned population inactive; and flat population inactive (Fig. 3B-E). To maintain the same average output firing rate of 5 Hz in the modulated conditions, we increased the activity of the remaining active inhibitory population (Fig. 3C).
When both populations were active, the input-output correlation remained unchanged (Fig. 3D, top), because the two populations (co-tuned and flat) were mimicking the effect of the single (co-tuned) population. Deactivating either population (while increasing the firing-rate of the other) had pronounced effects on postsynaptic responses. The average postsynaptic firing rate remained unchanged, but fluctuations in firing rate and membrane potential increased in both cases (Fig. 3B,C). When the co-tuned inhibitory population was turned off, the emerging imbalance of excitation and inhibition unmasked the excitatory tuning curve, increasing the chance of action potential generation when preferred signal populations were active (Fig. 3B, middle). The compensatory increase in the activity of the flat population further quenched non-preferred excitatory signals, leading to anti-correlated responses for non-preferred input signals (Fig. 3D, purple). The opposite effect could be observed when the flat population was deactivated. In this case, the lack of inhibition for non-preferred signals gave rise to input/output correlations for non-preferred signals, while preferred signals saw no response (Fig. 3B, right and Fig. 3D, green).
Compared to the control case (Fig. 3E, top), transient responses were drastically increased for preferred inputs when the co-tuned population was deactivated, and the response to non-preferred signals was completely diminished (Fig. 3E; middle). When the flat population was deactivated, the post-synaptic neuron responded strongly to the non-preferred inputs, but not to preferred inputs (Fig. 3E; bottom).
Co- and counter-tuned inhibitory populations
We also analysed the dynamics of two inhibitory populations which were tuned such that one population mirrored the tuning of the excitation (as before, co-tuned), while the other population was tuned such that the weakest excitatory signals received the strongest inhibition (Fig. 4A). We called this population counter-tuned. Counter-tuning has been observed as a consequence of anti-Hebbian inhibitory plasticity in experimental work in hippocampus21–23 and used in theoretical work on associative memory networks24. In the balanced state, output behaviour is almost identical to the scenarios described above (Fig. 4B-D; control). The main distinction between the models with counter-tuned or flat inhibitory profiles is how they complemented the co-tuned inhibitory currents: the flat inhibition produced currents that tracked the co-tuned current, whereas counter-tuned inhibition produced inhibitory currents that were largely uncorrelated to the co-tuned inhibitory currents (Fig. 4B, left; compare with Fig. 3B, left).
When either the co- or the counter-tuned inhibitory populations were inactivated, fluctuations in both firing rate and membrane potential increased considerably (Fig. 4B, middle and right). Deactivation of the co-tuned population resulted in positive correlation between postsynaptic activity and preferred signals, and negative correlation between output and non-preferred signals (Fig. 4D, purple).
For pulse stimulation, there was no discernible difference to the model with flat inhibition in the control state (Fig. 4E, top). Turning off counter-tuned inhibition (Fig. 4B-E) also had similar results in the postsynaptic response as turning off the flat inhibition (cf. Fig. 3B-E). Unlike before, turning off co-tuned inhibition produced elevated firing-rate responses also for transient stimuli from signals directly neighbouring the preferred input (Fig. 4E, middle row, compare with Fig. 3E, middle).
Quantitative differences of inhibitory profiles
For a quantitative evaluation of the differences between the three models with their varying inhibitory receptive fields, we introduced the parameter ΔC = 0.5(Cpref – Cnon-pref), i.e., 50% of the difference in input/output correlation between preferred, Cpref, and non-preferred, Cnon-pref, signals (Methods). In the control condition, we observed similar ΔC ≈ 0 in all models (Fig. 5A, grey), reflecting low levels of correlation between output and input signals (Fig. 5B, top). With down-regulated inhibition, ΔC increased slightly in the model with one homogeneous inhibitory population. ΔC increased more considerably in a two-population model in which the co-tuned population was inactive (Fig. 5A, purple), indicating an increased correlation between preferred signal and output (Fig. 5B, middle). When the flat or the counter-tuned inhibitory populations were in-activated, we observed postsynaptic responses even to non-preferred input signals (Fig. 5B, bottom), which led to negative ΔC. Inactivating the counter-tuned inhibition resulted in a slightly better discrimination of non-preferred input signals (Fig. 5A, green).
To quantify pulse responses, we considered how many pulse signals could be recovered across the receptive field (Fig. 5C). We counted only those responses with more than 50% of the maximum firing-rate (Fig. 5D). The single inhibitory population model could only produce responses to preferred input signals, while co-modulation of two inhibitory populations could promote responses to non-preferred input signals, as well. Counter-tuned population achieved better postsynaptic control than flat inhibition (Fig. 5C).
Plasticity shapes inhibitory receptive fields
We have shown that complementary tuning of two inhibitory populations allows the retrieval of various facets of mixed and coincident stimuli. We wondered how such tuning patterns could emerge from naïve connectivity. To study how plasticity can shape the emergence of opposite receptive fields, we incorporated inhibitory synaptic plasticity mechanisms into our model with two inhibitory populations. We first implemented a Hebbian rule (that potentiated synaptic weights for coincident pre- and postsynaptic spikes and depressed them for sole presynaptic spikes6, Fig. S1A), in one of the two inhibitory populations while the synapses of the excitatory and the other inhibitory population remained fixed (Fig. 6). This learning rule has previously been shown to generate inhibitory weight profiles that mirror the excitatory receptive field of a postsynaptic neuron, imposing a firing-rate fixing-point (target; Fig. S1B) by balancing excitation and inhibition. Simulations began with tuned excitatory synapses and flat inhibitory weight profiles in both inhibitory populations (Fig. 6A).
After 30 minutes of stimulation with natural inputs (cf. Fig. 1B), inhibitory weights of the plastic population stabilised (Fig. 6D-G). Whether the target firing rate (Fig. 6B,C) was reached depended on the synaptic strength of the other, static population of inhibitory synapses. If the static weights were weak, the plastic synapses increased their strength until the target firing rate was reached (Fig. 6C). If the static population provided strong inhibition (and thus kept postsynaptic firing below the target rate), weights from the plastic population would eventually vanish — before the target firing-rate could be reached (Fig. 6C,G).
The effective receptive field could be interpreted as the cumulative difference between excitatory and inhibitory weights. Consequently, the shape of the static population determined the shape of the plastic population (Fig. 6D,E). As expected, the input/output correlation of the postsynaptic responses followed the effective receptive field profile (Fig. 7A, cf. Fig. 6E). The best performance, i.e., distinct responses to preferred or non-preferred stimuli according to the modulatory state, and flat responses otherwise (Fig. 7B), emerged when the synapses of the static population were constrained to a narrow band of synaptic weights (Fig. 7C, shaded region).
Next, we introduced plasticity to the second population of inhibitory afferents. We tried two different rules. We began with a homeostatic plasticity rule (Methods) which scaled synapses to reach a fixed-point in the postsynaptic firing-rates (Fig. S1C), achieving results that matched the handed-tuned solution. Notably, this plasticity rule was purely local, taking only presynaptic weights and postsynaptic firing rates into account, similar to experimentally observed scaling of inhibitory synapses38,39. With the homeostatic rule co-active, the Hebbian synapses — connections changing according to the Hebbian plasticity rule — developed a co-tuned profile from initially random weights (Fig. 8A, top; Fig. 8C, left), while the synapses following the scaling rule collapsed to a single value (Fig. 8A, bottom; Fig. 8C, right). Consequently, the postsynaptic neuron received precisely balanced inputs (Fig. 8B), autonomously arriving in a regime of optimal performance (cf. Fig. 7C).
Instead of a purely homeostatic scaling rule, we also tried an experimentally observed21–23 anti-Hebbian rule in the second inhibitory population (Fig. S1D). The Hebbian plasticity rule in one synapse population leads to stable receptive fields and imposes a firing-rate set point for the postsynaptic neuron. The anti-Hebbian rule, on the other hand, increases the firing rate of the postsynaptic neuron indefinitely, because correlated activity decreases synaptic weights (only sole presynaptic spikes increase synaptic weights; Methods). Anti-Hebbian plasticity rules is thus unstable (Fig. S1E). To overcome this problem without incorporating additional, complex dynamics, we set the learning rate of the anti-Hebbian rule to decrease exponentially over time (Fig. S1F; see Discussion). With the anti-Hebbian rule active, initially random weights evolved into co-tuned and counter-tuned receptive field profiles (Fig. 8D). As learning slowed down, the anti-Hebbian synapses — connections changing according to the anti-Hebbian plasticity rule — stabilised, and Hebbian synapses ceased to change once the target firing rate was reached (Fig. 8E; Fig. 8F). The final shape of both receptive fields was nearly identical to the hand-tuned profiles (cf. Fig. 4A), with equivalent postsynaptic performances (data not shown; see also Fig. 3 and Fig. 4).
Discussion
We investigated how several distinctly tuned inhibitory and excitatory synaptic populations interact in a receptive field-like paradigm. In our model, we aimed for precise balance of excitation and inhibition, accordant with evidence of excitatory and inhibitory co-tuning in cat visual cortex40, rodent auditory cortex12,41,42 and rodent hippocampus43, and temporal correlations in neighbouring excitatory and inhibitory synapses44.
Consistent with earlier work, we could modulate the efficacy of a single inhibitory population to enhance the output correlation with the preferred input18,19, but the flexibility of the control mechanism was very limited. Non-preferred signals never evoked faithful responses (Fig. 2). We wondered if additional, independently tuned synapse populations would grant more flexible control over the signal stream.
We constructed models with two types of inhibitory interneurons that could be modulated independently and carried a diverse set of input signals. Each signal group of excitatory synapses had two corresponding signal groups of inhibitory synapses, all three displaying correlated activity patterns, with their combined efficacy balancing the net input current in the baseline (control) state. To emphasise different aspects of the overall input, we allowed for the ratio of inhibitory currents delivered by the two inhibitory populations to vary dynamically over time, consistent with evidence from macaque V1 indicating that the receptive fields of neurons are dynamic, rapidly switching their selectivity45. Our results suggest that a discrimination of preferred and non-preferred stimuli, such as occurs in selective attention, may be achieved by dynamically modulating inhibition. Such modulation of stimulus-selectivity could correspond to mechanisms of top-down selective attention46–48.
Attending to a specific stimulus is known to increase the firing rate of sensory cells selective for this stimulus49–51. Such sharpening of stimulus-selectivity was also observed in our simulations (Fig. 3 and Fig. 4; purple). We also observed the opposite: enhanced postsynaptic responses to non-preferred and suppressed responses to preferred stimuli (Fig. 3 and Fig. 4; green). We interpreted this as selectively ignoring a specific stimulus, rather than attending to it. Thus, the two aspects of selective attention – enhancing response to targets, and suppressing the response to distractors – were implemented in our model by the two types of disinhibition.
The architecture of our model fits well with the diversity of interneuron types and connectivity profiles in the neocortex1,4. Co-tuned inhibitory afferents may originate from parvalbumin-positive (PV+) interneurons. As the main source of inhibition to principal cells, PV+ interneurons tend to target postsynaptic neurons with similar preferred orientation52. Moreover, activation of these neurons leads to a flattened selectivity curve52 (but see Lee et al.53). Flat (or counter-tuned) inhibitory afferents may arrive from somatostatin-positive (SOM+) interneurons, with their less selective connectivity pattern52. This interpretation is in line with recent evidence suggesting that top-down visual attention relies on local inhibitory circuitry in primary visual cortex42. In this scheme, PV+ and SOM+ neurons inhibit pyramidal cells, while vasoactive intestinal peptide-positive (VIP+) neurons suppress other inhibitory interneurons, acting as a source of disinhibition. Direct manipulation of SOM+, PV+ and VIP+ neurons confirms these respective roles in inhibition and disinhibition in both visual54 and auditory cortices15,55. Additionally, Zhou et al.42 reported that VIP+ neurons received excitatory top-down inputs from the rodent cingulate cortex, allowing for a narrow selectivity profile of pyramidal cells when active and the opposite when inactive. Finally, blocking cortical inhibition reduces the stimulus-selectivity of cortical neurons56,57 (but see Nelson et al.58).
To show how these inhibitory synaptic profiles could develop, we implemented a Hebbian-like inhibitory plasticity rule6,20, co-active with either a homeostatic scaling38,39 (Fig. 8A-or an anti-Hebbian21–23 (Fig. 8D-F) plasticity rule. The scaling plasticity rule we implemented decreases synaptic weights proportionally to their current weight (multiplicative), and strengthens weights by a constant amount (additive), relative to the firing rate of the postsynaptic neuron. Consequently, all synapses tend to converge to the same value, without being enforced by a global rule. There is currently no direct evidence for this specific implementation, but it would provide a parsimonious explanation for the un-tuned, blanket inhibition often encountered in experiments59, and would allow modulating postsynaptic responses independently of the presynaptically tuned receptive field.
The anti-Hebbian rule we used had the opposite shape of the Hebbian plasticity rule, and was unstable, i.e., it could lead to infinite strengthening of presynaptic weights. A silent feedforward network would be the result of such connectivity (but see Hendin et al. 24). To obtain a plausible receptive field, we thus had to use an additional mechanism. The simplest solution was to include a modulatory term to the learning rate of the anti-Hebbian plasticity rule. This process mimics modulatory control of plasticity through the activity of other neuronal types12,60,61 and allowed us to balance the effect of the opposing rules. Other mechanisms could be used to control the unstable nature of the Anti-Hebbian plasticity rule, but the generality of the results would likely not be altered.
To explore the interaction between two distinct inhibitory plasticity rules without confounds, we made the simplifying assumption that the excitatory synapses would remain fixed (but see Clopath et al. 62, Litwin-Kumar and Doiron63, and Zenke et al.64). Further work could explore how inhibitory plasticity rules interact not only with each other, but also with rules for excitatory synaptic plasticity — thus getting closer to the realistic conditions of biological neural circuits. Additionally, incorporating more types of inhibitory interneurons is a possible next step to be explored, and thus the study of the interaction of more inhibitory plasticity rules9,65,66.
Our model only considered a single postsynaptic neuron, with no feedback connectivity, which is thought to be the source of most orientation-selective input to cortical cells67.
Further theoretical work should extend the combined use of these distinct plasticity rules for inhibitory synaptic connectivity to larger architectures, involving multiple neurons and incorporating feedback connections. Further synergies may emerge, such as multiplicative and additive modulation of receptive fields, and surround suppression68. Other possible functionalities that have been explored in this context are gating of different signal streams47,48, one-shot learning in hippocampus11, and consequences to brain disorders69.
Interestingly, artificial networks have been shown to develop similar receptive field profiles to the ones explored here when they are trained to solve multiple tasks70. Yang et al.70 have shown that clusters of neurons can acquire co-tuned or flat connectivity, which are controlled by context-encoding signals. These results hint at the possibility that biological and artificial systems may utilise similar strategies to solve context-dependent filtering tasks.
Finally, we predict that various GABAergic interneurons in the same cortical region must obey a range of different inhibitory synaptic plasticity rules, to restore or reverse neuronal stimulus-selectivity as appropriate and necessary. Such evidence would provide empirical validation to the theoretical results presented here, and in turn inform future computational modelling.
Materials and Methods
Detailed methods are given in the supplementary materials.
Software and code availability
Simulations were run in Fortran, compiled with Intel Fortran Compiler 19.0 on an Intel-based Linux computer (Debian 9; i9-9900X processor; 32 GB memory). Codes will be made available online upon publication71. Individual plots were generated with Gnuplot. Figures were generated with Inkscape.
Author contributions
EJA and TPV designed research; EJA and AIL carried out the simulations and analysis; EJA, AIL and TPV wrote the manuscript.
Supplementary material
Codes for all results are openly available at GitHub, repository https://github.com/ejagnes/attentional_switch.
Neuron model
To investigate changes in neuronal response due to specific inhibitory connectivity motif we simulated a postsynaptic leaky integrate-and-fire neuron (LIF) receiving excitatory and inhibitory afferents. Postsynaptic neuronal membrane potential dynamics is governed by where u(t) is the somatic voltage at time t, τm = RC is the membrane time constant (membrane resistance, R, times membrane conductance, C), urest is the resting membrane potential, and EE and EI are the reversal potential for excitatory and inhibitory synapses, respectively. Synaptic conductances, gE(t) and gI(t), evolve according to and Both excitatory and inhibitory conductances decay exponentially to zero with time constants τE and τI, respectively. Presynaptic action potentials trigger increase in synaptic conductances through the sum of Dirac delta functions, where tkj is the time of the kth spike of presynaptic afferent j. The contribution of a given presynaptic afferent j to changes in conductances is given by the synaptic weight wj(t), which was fixed for excitatory synapses and could change over time due to plasticity mechanisms for inhibitory synapses. The total number of presynaptic afferents is N = NE + NI, with NE being the number of excitatory and NI of inhibitory presynaptic afferents.
An action potential is triggered at the postsynaptic neuron once its membrane potential u(t) crosses the spiking threshold uth from below. The membrane potential is then instantaneously reset to ureset, being clamped at this value for the duration of the refractory period, τref. The postsynaptic spike train is described here as a sum of Dirac deltas, where tk is the time of the kth spike of the postsynaptic neuron, or the time the membrane potential crosses the spiking threshold from below. Parameters used for the postsynaptic neuron are detailed in Table I.
Inputs
To mimic receptive field input-like, we divided the synaptic inputs into P signal groups (µ = 1, …, P) that share the same fluctuation in firing rate. We tested two cases: (i), natural input, and (ii), pulse input. Both are described below.
Natural input
For presynaptic activity mimicking a natural input, activity follows an inhomogeneous Poisson process that changes according to a modified Ornstein-Uhlenbeck (OU) process. We first defined an auxiliary variable for each pattern, yµ(t), that follows a stochastic first-order differential equation given by where µ is the signal group index, τOU is the time constant for the decaying process that changes due to a Gaussian noise term ξµ(t) with unitary standard deviation. The mean value of the variable yµ is zero, and thus it assumes positive and negative values with same probability (for long periods).
The spike train of an afferent in a given signal group µ is given by the variable νXµ(t) which is a rectified version of the auxiliary variable plus a term to generate background firing rate, νXbg, where X indicates the presynaptic population; X = E for excitatory and X = I for inhibitory. The spike trains of the afferents of signal group µ are generated by where νX0 is the amplitude of the modulated firing rate fluctuations, and [·]+ is a rectifying function, Note that due to the symmetry of yµ(t), an afferent is half the time in the background state and half the time in the active state.
Presynaptic action potentials were generated as an inhomogenous Poisson process according to the modified OU process described above and a fixed background firing rate. Additionally, we implemented a refractory period, τEref for excitatory and τIref for inhibitory inputs. Given the time step of the simulation Δt, spikes of a presynaptic afferent that is part of the signal group µ are generated with a probability pXµ(t) = νXµ(t)Δt if there was no spike elicited during the refractory period before-hand, and thus the average firing rate of a X = E (excitatory) or X = I (inhibitory) afferent that is part of the signal group µ becomes
Pulse input
To test transient responses to brief changes in presynaptic activity we also quantified postsynaptic responses to pulse inputs. In this case, we simulated the postsynaptic neuron receiving inputs with constant background firing-rate. For 100 ms we increased the probability of presynaptic spikes for a given signal group µ by a factor kν*, with k being an integer larger or equal than zero, and ν* = 5 Hz. Thus presynaptic spikes are generated by during the 100 ms step and by during only background activity. Parameter αXν is a scalar that sets the ratio of excitatory and inhibitory firing rate.
Responses to the pulse input were divided in two bins: phasic and tonic. Phasic responses were defined as the postsynaptic activity elicited in the first 50 ms of the pulse input. Tonic activity was correspondingly defined as having occurred in the last 50 ms of the stimulus. We simulated 100 trials per input strength kν*, and defined the response (for both phasic and tonic) as the average number of spikes on the period for the strength kν* minus the average number of spikes on the same period without extra input, multiplied by 20 to convert to Hz. We subtracted background spikes to ascertain that we quantified the response to the extra step input alone. Parameters used for inputs are detailed in Table II.
Synaptic tuning
Based on Vogels et al.6, we used a receptive field profile given by where r0, b and c are parameters defining the shape of the receptive field and µ0 defines the preferred signal group, which maximises r(µ); r(µ0) = 1. Note that r0 ≥ 1, b ≤ 1, µ0 > 0, and c is an even positive integer.
For simplicity we define ζj as the signal group that afferent j is part of. Thus excitatory synapses are set as where wE0 is a normalisation factor for excitatory weights, and ϵj is a noise term drawn from a uniform random distribution between and .
First we simulated a single inhibitory population with a tuned profile (Fig. 2), following Eq. 12 such that, where wI0 is a normalisation factor for inhibitory weights, and ϵj is a noise term drawn from a uniform random distribution between and . The parameter wI0 was chosen so that a state of balance was enforced, with postsynaptic firing-rate of 5 Hz. Due to the small number of inhibitory afferents compared to the excitatory ones, and the difference in driving force, inhibitory weights were much larger than excitatory ones. Thus, to plot excitatory and inhibitory weights on the same scale we computed the correcting factor, αw, from In all plots with excitatory and inhibitory weights, we plotted excitatory weights multiplied by the parameter αw.
Next, inhibitory afferents were divided in two types that connect to the postsynaptic neuron following different receptive field profiles. We used two approaches to sculpt the inhibitory synaptic weight profiles. In our first approach we defined them based on Eq. 12, i.e., we hand-tuned their shape. For our second approach we started them randomly and applied plasticity rules so that they produced the previously hand-tuned shapes through unsupervised learning.
We combined a co-tuned population with either a flat (Fig. 3) or a counter-tuned (Fig. 4) population. In both cases we based co-tuned weights on a modified version of Eq. 12, with b, µ0 and c the same as for Eq. 12. Inhibitory weights for the co-tuned profiles are chosen so that where wIco is a normalisation factor for inhibitory weights following the co-tuned receptive profile, and are different when combined with either the flat or the counter-tuned populations.
We set the flat population such that where wIf is the average for the flat population. The shape of the counter-tuned population was defined by and synapses were hence tuned such that where [·]+ is a rectifier (Eq. 8), used to enforce only positive synaptic weights. The parameters wIcounter is a normalisation factor for the counter-tuned inhibitory populations.
When plasticity was simulated, initial conditions for all plastic inhibitory populations were flat with noise (Fig. 6, Fig. 7 and Fig. 8), Parameters used for the tuning curves are detailed in Table II, and for synaptic weights in Table III. Both the average of the weights for flat population, wIf, and the noise term, , were distinct for different simulations.
Plasticity models
In this work we used three different inhibitory synaptic plasticity (ISP) rules. We termed them Hebbian, scaling, and anti-Hebbian. Both Hebbian and anti-Hebbian plasticity rules are triggered by pre- and postsynaptic spikes, and depend on a low-pass filter of these spike trains. The presynaptic trace (low-pass filter) is given by where xj(t) is the value of the trace of the spike train of presynaptic afferent j at time t; τSTDP is the time constant of the trace, and Sj(t) is a sum of Dirac delta functions (Eq. 4) representing the spike train of afferent j. The same is considered for the postsynaptic neuron, where xpost(t) is the postsynaptic trace, and Spost(t) is the spike train of the postsynaptic neuron (Eq. 5). Note that we used the same time constant for both pre- and postsynaptic traces.
Hebbian inhibitory plasticity
Precise balance of excitatory and inhibitory inputs was learned by a Hebbian inhibitory plasticity rule6. The weight of the jth inhibitory synapse changes according to where ηH is the learning rate, and αH is a parameter that defines the postsynaptic firing-rate. The first two terms on the right-hand side of Eq. 24 are Hebbian terms that increase the weights when both pre- and postsynaptic activities are correlated. The last term on the right-hand side of Eq. 24 is a penalty term for inhibitory spikes alone, which creates a homeostatic set-point for the postsynaptic firing-rate given by Later we describe how to arrive at this approximation.
Inhibitory synaptic scaling for flat tuning
One of the receptive profiles we used for inhibitory synapses was flat, i.e., every synapse group had the same strength. To learn the flat pro-file from random initial weights we implemented a scaling plasticity rule, partially based on experimental work that observed synaptic scaling on inhibitory synapses38,39. Weights are increased if the postsynaptic firing-rates are too high, and decreased otherwise, where ηs is a learning rate, ρ0 is a firing-rate reference value, chosen to be the same as the one for Hebbian plasticity rule, Θ(·) is the Heaviside function and αs is a term that sets the firing-rate range for which synapses do not change. Postsynaptic neuron’s firing-rate is computed with a slow averaging of the postsynaptic activity through where τscaling is the time constant for the postsynaptic activity and Spost(t) is the postsynaptic spike train (Eq. 5). Note that the last term on the right-hand side of the equation above is divided by τscaling so that ypost(t) is in units of rate. Synaptic depression is weight dependent while synaptic potentiation is not, which ensures that all synaptic weights tend to the same value. When the postsynaptic neuron is firing below a threshold ρ0/αs, all inhibitory synapses in the flat group have their weights decreased proportionally to the difference between the target firing-rate and the average firing-rate, but also proportional to the current weight value. This way, strong synapses undergo stronger decrease than weak ones. Conversely, when the postsynaptic neuron is firing above a threshold αsρ0, the same synapses increase in value by the same amount. Intuitively, these mechanism ensures that all synapses converge to the same value for a long run.
Anti-Hebbian inhibitory plasticity
The third inhibitory plasticity rule we used is an anti-Hebbian rule based on experimental data21–23 and theoretical work on recurrent networks24. Synaptic weights change according to where ηaH(t) is a variable learning rate and αaH is a parameter to counterbalance the anti-Hebbian term. The resulting rule dictates that coincident events decrease inhibitory synapses, while non-coincident ones increase synaptic weights. Due to the unstable nature of this plasticity rule (see details below), we implemented a time-varying learning rate which evolves according to where τaH is the decay time constant for the learning rate, and MaH(t) is an external signal to transiently activate plasticity. We speculate that such signal could come from modulatory neurons such as dopaminergic or cholinergic and assumed that the external signal peaks at a time t0 (beginning of the simulation), so that where is the maximum learning rate before decaying to zero, and t0 is the time when plasticity at these synapses are initiated. Parameters used for plasticity models are detailed in Table IV.
Mean-field analysis of the plasticity rules
We were interested in plasticity rules with stable dynamics. For a better intuition on fixed-point dynamics and stability we consider here a simplified dynamics of a mean-field model for both the Hebbian6 and the anti-Hebbian models. We define the post-synaptic firing-rate as νpost(t) and the presynaptic firing-rates as νj(t). The traces of both presynaptic afferent and postsynaptic neuron thus have an average of τSTDP νj(t) and τSTDPνpost(t), respectively64. Neglecting any correlation between pre- and postsynaptic spikes, the average weight change for Hebbian synapses is given by where ⟨·⟩ represents average over time. Intuitively, the post-synaptic firing-rate, νpost(t), changes negatively with changes in inhibitory weights — increased inhibition generates fewer postsynaptic spikes and vice-versa for decreased inhibition. This means that average firing rates are inversely linked to average inhibitory weights, i.e., The steady state is computed by considering the vanishing point of the equation above (we assume that the presynaptic activity is nonzero), thus This means that the postsynaptic activity νpost(t) increases (via decreases in inhibitory efficacy) when below ρ0 and decreases when above ρ0, creating a stable fixed-point for the postsynaptic firing-rate.
The opposite is true for the anti-Hebbian plasticity rule. Changes in postsynaptic firing-rate (with the same assumption as for the Hebbian plasticity rule) follow Because postsynaptic activity increases when it is above threshold ρ1 and decreases when it is below, this rule is un-stable. The postsynaptic firing-rate eventually explodes or vanishes. We chose the simplest way to overcome these problems by setting a time-varying learning-rate. Other intricate mechanisms could be implemented, but this is not the scope of our work
Correlation
We quantified the response of the postsynaptic neuron to natural inputs with the Pearson correlation between postsynaptic firing-rate and input firing-rate fluctuations, per signal group. We computed the firing rate of a signal groups as the low-pass filter of the spike trains of its excitatory afferents, where Zµ(t) is the firing rate of the signal group µ at time t, filtered with a time constant τZ. The postsynaptic activity is also computed through a low-pass filter of its spike train, where ϒ(t) is the activity of the postsynaptic neuron at time t filtered with a time constant τϒ. The correlation is then computed as where cov(z, y) is the covariance between variables z and y, σz is the standard deviation of variable z, and ⟨·⟩ represents time average.
Subsequently we computed a performance index ΔC as the difference between the correlation measure for preferred (µ =9) and non-preferred (µ = 1) input signals, Maximum positive performance index, ΔC = 1, means that the preferred signal group has maximum correlation (C9 = 1) while the non-preferred signal group has maximum anti-correlation (C1 = −1), indicating that the postsynaptic neuron is responding solely to the preferred signal group. Consequently, ΔC = −1, indicates that the postsynaptic neuron is responding solely to the non-preferred signal group. A flat response is indicated by ΔC = 0. Note that maximum ΔC (either positive and negative) is only achievable if there is no overlap between activation of preferred and non-preferred input signals, which is never the case here. We define as best performance when ΔC = 0 for all inhibitory inputs active (control), ΔC = 1 (or IµΔC > 0) for one inhibitory population inactive, and ΔC = −1 (or ΔC < 0) when the other inhibitory population is inactive. Parameters used for computing correlations are detailed in Table V.
Implementation
Models were simulated with a time-step Δt, with either analytical or semi-analytical solution of the corresponding differential equation. All codes were written in Fortran, compiled with Intel Fortran Compiler 19.0, running on an Intel-based Linux computer (Debian 9; i9-9900X processor; 32 GB memory). Bellow we describe how each equation was implemented, with the parameter values in tables in the end of this section.
When not in the refractory period (see below), the leaky integrate-and-fire neuron is updated as where n is the iteration index, and are auxiliary variables described by and This is the analytical solution when considering that all variables apart from u(t) are constant during a time-step, which we refer to as semi-analytical.
When the membrane potential crosses a threshold from below, the membrane potential is reset (because of a spike being triggered), and kept at the reset potential for the duration of the refractory period, where Synaptic conductances are implemented as
Note that here is equal to one when afferent j spiked at time-step n and zero otherwise.
For natural inputs, we updated the auxiliary variable yµ(t) every 1 millisecond, where mod(·, ·) is the modulo operation and is a random number drawn from a gaussian distribution with zero mean and unitary standard deviation. Presynaptic spikes are generated as point processes, so that at each time-step the probability of a presynaptic afferent to spike is and during the τEref/Δt iterations after a spike. The same is valid for an inhibitory afferent; the probability of firing an action potential is and during the τIref/Δt iterations after a spike.
For pulse inputs, presynaptic afferents were set to fire at background firing-rate and had an elevated firing-rate during a 100-ms period, which was varied in 5 Hz steps. For the activated pattern where αEν adjusts the excitatory firing-rate, k is an integer for varying the pulse intensity, and n0 is the iteration in which the pulse starts. The same implementation was used for inhibitory afferents ( being the parameter to adjust the inhibitory firing-rate), An afferent in an inactive pattern fires action potentials with background frequency ( and ), and there is no spike elicit in the refractory period ( and during the τEref/Δt and τIref/Δt iterations after a spike, respectively).
Plasticity was implemented with spike triggered events. For the Hebbian and anti-Hebbian plasticity rules, auxiliary variables changed as where if the postsynaptic neurons generated an action potential at iteration n and zero otherwise. Hebbian weights changed according to and anti-Hebbian to with the learning rate varying as With , and for n > 0.
Scaling was implemented with a different trace, with weight update following Correlation-related variables were updated as and
Acknowledgements
We thank the members of the Vogels group, and particularly Georgia Christodoulou and William Podlaski, for fruitful discussions. This work was supported by a Research Project Grant by the Leverhulme Trust (RPG-2016-446; EJA), a Sir Henry Dale Fellowship by the Wellcome Trust and the Royal Society (WT100000; EJA, TPV), a Wellcome Trust Senior Research Fellowship (214316/Z/18/Z; EJA, TPV), and a Claredon Scholarship from the University of Oxford (AIL).
Rerences
- [1].↵
- [2].↵
- [3].
- [4].↵
- [5].↵
- [6].↵
- [7].↵
- [8].
- [9].↵
- [10].
- [11].↵
- [12].↵
- [13].↵
- [14].↵
- [15].↵
- [16].↵
- [17].↵
- [18].↵
- [19].↵
- [20].↵
- [21].↵
- [22].
- [23].↵
- [24].↵
- [25].↵
- [26].↵
- [27].↵
- [28].↵
- [29].↵
- [30].↵
- [31].↵
- [32].↵
- [33].↵
- [34].↵
- [35].↵
- [36].
- [37].↵
- [38].↵
- [39].↵
- [40].↵
- [41].↵
- [42].↵
- [43].↵
- [44].↵
- [45].↵
- [46].↵
- [47].↵
- [48].↵
- [49].↵
- [50].
- [51].↵
- [52].↵
- [53].↵
- [54].↵
- [55].↵
- [56].↵
- [57].↵
- [58].↵
- [59].↵
- [60].↵
- [61].↵
- [62].↵
- [63].↵
- [64].↵
- [65].↵
- [66].↵
- [67].↵
- [68].↵
- [69].↵
- [70].↵
- [71].↵