The basic understanding of the dynamical behavior of Spiking Neural Networks (SNN) in Neuroscience is the focus of intense research. There are many relevant and pressing questions that are coming into focus, for instance, breaking the neural code what is the nature of the neural code? how information is encoded and transmitted with spikes from one part of the brain to another? how does that depend on network topology? are brain networks close to a critical or a chaotic state? what is the robustness of networks to chaos? how neurons may synchronize to form waves? how dynamical memories are realized and sustained? and many others (Korn and Faure, 2003; Rabinovich et al., 2006). These issues are being studied either by in vivo and in vitro experiments in neurobiology (Reyes, 2003; Yin et al., 2018), or theoretically by means of numerical simulations of mathematical models of neural networks (Izhikevich and Edelman, 2008; Stimberg et al., 2019). In the first case, the neurobiological experiments are technically challenging and, evidently, one cannot systematically modify the neural networks. In the second, the results on mathematical modeling may always be questioned, as a peculiar result that may depend on the assumptions made. For instance, it may be hard to assess if the relative ubiquity of chaotic behavior (Korn and Faure, 2003; Rudolph and Destexhe, 2007; Nobukawa et al., 2018) is of biological relevance. Moreover, in a well-known study, Izhikevich and Edelman (2008) observed the surprising result that the suppression of a single neuron out of an ensemble of a thousand spiking neurons may change the state of the entire network.

In order to better position our current work with respect to other approaches, we should also briefly describe the current efforts on applied research in Neuromorphic Computational Electronic Engineering for artificial intelligence (AI). That field can be roughly divided in two big areas. One aims to implement dedicated computer processors, which are optimized to run algorithms based on mathematical models of neurons. Examples of those digital neuron chips TrueNorth, Loihi, SpiNNaker, etc (Furber, 2016; Thakur et al., 2018). These systems are multi-core chips with neuromorphic architecture, that is, they have vast numbers of relatively small memory and processing units, which are densely interconnected. We may also include in this category the recent implementations using Field Programmable Gate Arrays (FPGA), which are making fast progress (Yang et al., 2015, 2018, 2019, 2020). On the other hand, a qualitatively different electronic engineering approach is aimed to design CMOS VLSI circuits, called silicon neurons, which implement the neuron models directly in hardware (Indiveri et al., 2011). Those neurons can then be interconnected off-chip to form networks by means, for instance, of the address event representation (AER) (Liu et al., 2015). In this approach, no actual spikes are transmitted between neurons, but it is the information of a firing event which is sent using the AER protocol between neuron addresses. One prominent example of a CMOS VLSI system is the BrainScaleS chip, which implement AdEx neurons (Schemmel et al., 2010) and further examples are discussed in Furber (2016); Thakur et al. (2018). In those approaches, the goal is to implement extremely large numbers of neurons (and synapses) to achieve the computing capacity of a brain. Some of their most significant challenges are to achieve low power dissipation and miniaturization. To reduce the power, one may work with transistors in the subthreshold regime (Mahowald and Douglas, 1991; Benjamin et al., 2014), however, additional issues arise in that case, such as device variability (Indiveri et al., 2011). Regarding the neuron circuit miniaturization, reducing its physical dimensions remains a challenge due to the relative large size of the capacitor that is required to represent the voltage of the membrane.

The limitations of silicon neurons may be potentially overcome by exploiting the neuromorphic functionalities of quantum materials (del Valle et al., 2018). For instance, phase change materials (Tuma et al., 2016), which accumulate phase instead of charge. Another notable example are the Mott insulators, whose memristive properties permit the firing of an electric spike when they are driven across the insulator to metal transition (Janod et al., 2016). Despite intense activity, this field of research is emerging and currently proposed devices are still individual single neurons (Stoliar et al., 2017; Yi et al., 2018). The main challenges ahead are to achieve a reliable control of the materials and the understanding of their fundamental physical behavior (del Valle et al., 2019), before actual circuits may be implemented.

In the context of the approaches that deal with SNN systems that we described above, our present methodology shares features of many but is qualitatively different to all of them. The main goal of the present work is to bring to the research community a novel way to build and study SNNs of unprecedented simplicity, which opens a new way to do experimental work in basic neuroscience. We propose a methodology to build general neural circuits of a priori arbitrary topology, where spiking neurons are in direct interaction and are interconnected as lego-like blocks. To illustrate this point, we shall adopt a classic model of neuroscience: the Jeffress model of binaural detection of sound directionality (Jeffress, 1948).

Our methodology is built around a recently introduced electronic neuron circuit, the ultra-compact neuron (UCN) (Rozenberg et al., 2019). This UCN, like Mott neurons exploits a memristive behavior, however, in contrast to those it is not based on a quantum material but on a conventional electronics device, the thyristor. So in this regard, the UCN may be considered a silicon neuron model. The silicon neurons vary in their degree of circuit complexity, for instance realizations of the AdEx model may require tens of transistors (Indiveri et al., 2011). The UCN, in contrast, counts with a small number of components so it can be considered a compact neuron model, following the terminology introduced by Indiveri et al. (2011). Moreover, the UCN can be termed ultra-compact as it requires a minimal number of components. In fact, to realize a basic circuit of a leaky-integrate-and-fire (LIF) neuron model, the UCN requires, like most other silicon neurons, a capacitor to integrate charge, and a resistor to mimic the leaky feature. However, the novelty of the UCN is that the fire functionality can be realized by a single silicon controlled rectifier (SCR or thyristor), which is a conventional electronic device introduced in 1956. The key insight is to realize that the I-V characteristics of the SCR display memristive features analogous to that of Mott materials, which enable the neuromorphic electric spiking behavior. However, the SCR present the advantage that they are available off-the-shelf to implement artificial neurons (Rozenberg et al., 2019), avoiding the need to deal with the complexities of Mott quantum material fabrication and control. We should also mention that the UCN is not restricted to the LIF model as it can be easily extended to realize a large variety of biologically relevant spiking neuron models (Stoliar et al., 2020).

We should make clear that the present UCN based electronic circuit methodology is not aimed at competing with engineering implementations for AI, therefore, speed, dissipation, and physical size is not our immediate concern. Here, we explicitly demonstrate that one can make a crucial step up and go from the level of a single UCN spiking, to a functional and biologically relevant circuit of more than ten neurons. One qualitative difference with respect to neurocomputing engineering implementations is that the UCN network operates in real continuous time, with time scales in principle compatible with those of biological systems. The aim of our work is a proof-of-concept for a novel, flexible, and affordable platform to construct spiking neural networks. This methodology can be further scaled up to hundreds or possibly thousands of neurons, which can be implemented with low-cost printed circuits boards (PCB). This order of magnitude in the number of neurons is sufficient to study questions of basic neuroscience, such as those that we mentioned in the beginning. In other words, we aim at opening a new experimental field of basic research in Spiking Neural Networks to a potentially large community, which is at the crossroads of neurobiology, dynamical systems, theoretical neuroscience, condensed matter physics, neuromorphic engineering, artificial intelligence, and complex systems.

The paper is organized as follows: We shall first describe the basic features of the UCN circuit and of the Jeffress model SNN. We shall then show how UCN blocks can be used to implement the network, proceeding in two steps: Firstly, we shall construct a novel long-axon neuron, which will serve as the delay-line of the Jeffress model. Secondly, we shall combine two such long-axon neurons with a layer of output neurons that detect the coincidence of spikes propagating along the delay lines. We shall then demonstrate the neuro-computational functionality of sound directionality detection. Finally, we discuss the prospects of adopting the UCN as a general and easy to implement spiking neuron lego-like block for experimental electronic neuroscience and neuro-engineering.

We now turn to describe the behavior of our neuromorphic model. We begin with the long-axon neuron. In Figure 5 we show the experimental data of the spike propagation along the axon. We observe that a spike travels down the axon as each pulse at a given segment unit induces a subsequent pulse at its downstream neighbor. For the chosen values of the circuit components (see Supplementary Material) the pulse moves along the axon at a constant rate of approximately 0.045 segment/μs. This rate may be further increased by decreasing the capacitance of the UCN.

We may now present the experimental data for the full Jeffress model implementation. This is shown in Figure 6, where we show two examples of the system’s behavior upon arrival of an input to the left and to right ears, with a given ITD. We shall begin with the case of no delay between arrivals, i.e., ITD = 0. This is shown in the Figure 6A. We can observe that the action potentials propagate through each long-axon neuron in opposite directions and at the same speed (first and second panels from the top). Indeed, they arrive to the last segment of each long-axon at about the same time 80 μs. The small differences between the timings and propagations between the two long-axon neurons is due to the small variability among electronic components, which is specified by the maker. An important point to make here is that the representation of time in SNNs is a non-trivial task (Liu et al., 2015). It usually requires off-chip communication and an AER protocol. In contrast, our circuit implementation is operating directly in real continuous-time. Moreover, the time-scales of our model are compatible with biological neuron networks ones.

As shown in the diagram of Figure 4, we see that the detection of coincidences is conducted between the UCN pairs L3-R1, L2-R2, and L1-R3. From the data in Figure 6A we can see that the “crossing” of the propagating spikes is, as expected, between the second segment of each long axon, namely, the pair L2-R2. The traces of the membrane potentials of the output neurons is shown in the third panel from the top of Figure 6A. We see that the corresponding “center” neuron senses the arrival of the pulses in coincidence and consequently increases its potential that overcomes the threshold of ∼ 2 V and fires. The fire event of the output neuron OC that signals the detection of sound from the central quadrant can be seen in the bottom panels of the Figure 6A. We see a strong spike in the trace of OC, however, another smaller one can also be observed for OR. These latter one can be considered a residual effect that are rather natural to this system’s bio-mimetic neural architecture. We will come back to this point later on.

In the panels of Figure 6B we show the case of a finite delay arriving to the L and R ear input neurons. As can be seen in the top panels, the approximate delay is of 40 μs, arriving earlier at the “left” ear L0, which indicates that the sound would have come from the left sector.

From the propagation of the spikes depicted in these panels, we observe that coincidence is occurring between UCN pair L3-R1, which is detected by the output neuron OL. In this case, the third panel shows the membrane potential of “center” and “right,” and we see that they detect the passing of spikes through the long-axons but not in coincidence. Hence their voltage initially increases and then starts to leak. By the time the second pulse is detected, although the V_C has not yet fully relaxed, the new increase is not enough to reach the threshold. The V_C fall just short of overcoming the 2 V value. Thus, the sole coincidence is detected by the “left” output neuron OL, as can be seen in from its firing event in the bottom panel of the Figure 6B.

Comparing the input spikes and the output fire, we may notice the existence of a small delay, which is indicated in the traces of the output neurons OC in Figure 6A and of OL in Figure 6B with small arrows. The delay is of a few microseconds and is simply due to the time it takes for the output neuron to integrate the coincident input signals at its dendrites and fire, namely, to perform its neurocomputation.

In Figure 7 we finally show the systematic behavior of the output as a function of the time difference between the arrival of pulses to the left and right “ears” (L0 and R0) or ITD. We observe and approximately left-right symmetric behavior of the three output neurons with respect to the arrival delay. The output is computed and coded in terms of the pulse width of the output signal, as was indicated before in the case of the central sound incidence, or ITD = 0, shown in Figure 6A. From this type of coding we observe that similarly as in biological systems, there is a continuum of detection and we adopt the strongest signal (winner takes all) as the result of the neurocomputation.

As it can also be seen in Figure 7, the resolution of the time difference of our system is about 40 μs, which corresponds an azimuthal angular of about 4°, assuming an ear-to-ear distance of 0.2 m and 343 m/s for the speed of sound (Glackin et al., 2010; see Supplementary Material). In our present simple implementation, the angular resolution depends on the time it takes for a spike to propagate along two subsequent Ranvier nodes of the long-axon neuron (see Figure 5 and Supplementary Material). This timescale depends on the specific values of the UCN parameters. For instance, a bigger capacitor would require more time to get charged up to the fire threshold value, thus would increase the time it takes the spike to travel down the long axon. Hence this would decrease the angular resolution of the system (see Supplementary Material). Conversely, the resolution can be systematically improved by performing faster integration, which can be achieved using smaller capacitors for the membrane potential. Therefore the resolution depends on the specific values of the circuit components of the UCN blocks of the long-axon neuron (Rozenberg et al., 2019). Note, however, that given a delay resolution, the maximal angle that the model may detect depends on the number of Ranvier nodes implemented in the long-axon. In the Supplementary Material we provide details on how to implement a large sound detection angle.

Finally, from the data of Figure 7 we may also notice that the symmetry of the response is affected by a small off-set shift of about 10 μs. This is simply due to the fact that the SCRs, as all electronic components, have a small dispersion, or tolerance, in their production. This feature can be easily corrected by a feed-back loop of the output, which may tune one of the resistors at the gate of the SCR that sets the firing threshold. This correction can be cast in terms of a supervised learning of the auditory system, where a correction pulse generated by the “wrong” output may tune the value of a memristor at the gate of the SCR. Details of that procedure are described in the Supplementary Material.

We have shown how a functional spiking neural network of biological relevance and be constructed using the recently introduced ultra-compact neurons as building lego-like blocks. The specific neuronal network is the Jeffress model of the auditory system of animals and humans, which remains a central paradigm in the field more than 70 years after its original proposal. This model was also adopted by Lazzaro and Mead (1989) in their pioneering work to demonstrate the concept of neurocomputing.

To implement the Jeffress model, we have first introduced a novel long-axon neuron, which we motivated by analogy with the Hodgkin-Huxley model for axon propagation of action potentials. Our long axon-neurons were built as a chain of UCN units, where the initial one plays the role of a soma + axon-hillock. The spikes then propagate along a chain of UCN segments forming a transmission line of action potentials, analogous to the nodes of Ranvier of biological axons. Two of such long-axon units constitute the input neurons that run antiparallel to each other. They are interconnected via an output layer of two-dendritic inputs neurons, which are also implemented by UCN blocks. The result of the neuro-computation is coded by the spike intensity of the output neurons.

The present work is a concrete demonstration of how the UCN is a simple and versatile (Hodgkin and Huxley, 1952) building block that allows to construct bio-inspired spiking neural networks. This simplicity allows the UCN networks to be readily made from off-the-shelf electronic components. We should emphasize the important feature that the UCN is a physical model of a spiking artificial neuron that does not include unphysical features such a hard voltage reset nor an abstract spike signal. Those artificial characteristics are indeed present in the integrate and fire mathematical models and are a source of concern for the stability of the dynamics (Korn and Faure, 2003; Nobukawa et al., 2018).

One may envision that neuroscientists may exploit the simplicity and convenience of the present scheme in multiple manners. For instance, they may begin to build physical SNN of UCNs to address under a new light open issues, such as the nature of the neural code in feed-forward and recurrent neural networks. So far, the study of these important questions has essentially been restricted to numerical simulations of schematic mathematical models and algorithms.

The present work introduces a novel experimental platform for the systematic construction and study of the dynamical behavior of SNN. This platform is simple, versatile and affordable, so it may open a new avenue of research in experimental neurobiology at a large and exciting interdisciplinary crossroad.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.