Synaptic Integration of Spatiotemporal Features with a Dynamic Neuromorphic Processor

Spiking neural networks have shown great promise for the design of low-power sensory-processing and edge-computing hardware platforms. However, implementing onchip learning algorithms on such architectures is still an open challenge, especially for multi-layer networks that rely on the back-propagation algorithm. In this paper, we present a spike-based learning method that approximates back-propagation using local weight update mechanisms and which is compatible with mixed-signal analog/digital neuromorphic circuits. We introduce a network architecture that enables synaptic weight update mechanisms to back-propagate error signals across layers and present a network that can be trained to distinguish between two spike-based patterns that have identical mean firing rates, but different spike-timings. This work represents a first step towards the design of ultra-low power mixed-signal neuromorphic processing systems with on-chip learning circuits that can be trained to recognize different spatio-temporal patterns of spiking activity (e.g. produced by event-based vision or auditory sensors).

Intracellular study of the cat's primary auditory cortex

1. We recorded intracellularly from X and Y cells of the cat's lateral geniculate nucleus and measured the postsynaptic potentials (PSPs) evoked from electrical stimulation of the optic chiasm. We used an in vivo preparation and computer averaged the PSPs to enhance their signal-to-noise ratio. 2. The vast majority (46 of 50) of our sample of X and Y cells responded to stimulation of the optic chiasm with an excitatory postsynaptic potential (EPSP) followed by an inhibitory postsynaptic potential (IPSP); these were tentatively identified as relay cells. We quantified several parameters of these PSPs, including amplitude, latency, time to peak (i.e., rise time), and duration. 3. Among the relay cells, the latencies of both the EPSP and action potential evoked by optic chiasm stimulation were shorter in Y cells than in X cells. Furthermore, the difference between the latencies of the EPSP and action potential was shorter for Y cells than for X cells. This means that the EPSPs generated in Y cells reached threshold for generation of action potentials faster than did those in X cells. The EPSPs of Y cells also displayed larger amplitudes and faster rise times than did those in X cells, but neither of these distinctions was sufficient to explain the shorter latency difference between the EPSP and action potential for Y cells. 4. The EPSPs recorded in relay Y cells had longer durations than did those in relay X cells. Our data suggest that the subsequent IPSP actively terminates the EPSP, which, in turn, suggests that the time interval between EPSP and IPSP onsets is longer in Y cells than in X cells. Furthermore, we found that, for individual Y cells, the latency and duration of the evoked EPSP were inversely related. These observations lead to the conclusion that the latency of IPSPs activated from the optic chiasm is relatively constant among Y cells and thus independent of the EPSP latencies. Thus the excitation and inhibition produced in individual geniculate Y cells may originate from different populations of retinogeniculate axons. 5. The IPSPs recorded in geniculate relay cells following optic chiasm stimulation could be divided into three groups based on their durations. The majority of both X and Y cells showed short-duration IPSPs, whereas the remainder of Y cells displayed medium-duration IPSPs, and the remaining X cells displayed long-duration IPSPs. A positive correlation was seen between the time to peak and duration of these IPSPs.(ABSTRACT TRUNCATED AT 400 WORDS)

Neuromorphic speech recognition systems that use spiking neural networks (SNNs) and memristors are progressing in hardware development. The conventional manual preprocessing of audio signals is shifting toward event-based recognition with convolutional SNNs. Despite achieving high accuracy in classification, the efficient extraction of spatiotemporal features from audio events continues to be a substantial challenge. In this study, we introduce dynamic time-surface neurons (DTSNs) using volatile memristors featuring an adjustable temporal kernel decay, enabled by series-connected transistors with an Au/LiCoO 2 /Au configuration. DTSNs act as feature descriptors, enhancing the spatiotemporal feature extraction from event audio data. A two-layer SNN classifier, fully connected and incorporating a 1T1R nonvolatile memristor array, is trained to recognize the spatiotemporal features of the audio data. Our findings show classification accuracies of up to 95.91%, substantial improvements in computational efficiency, and increased noise resilience, confirming the promise of our memristor-based speech recognition system for practical applications.

Conditions required for polysynaptic excitation of dentate granule cells by area CA3 pyramidal cells in rat hippocampal slices.

1. The recruitment of evoked fast inhibitory postsynaptic currents (IPSCs) and excitatory postsynaptic currents (EPSCs) was examined using whole cell voltage-clamp recordings from layer V pyramidal neurons in slices of rat somatosensory cortex. Synaptic currents were evoked with graded electrical stimulation to assess the relative activation of IPSCs and EPSCs. Fast GABAA ergic IPSCs were selectively recorded by holding cells at potentials equal to EPSC reversal (approximately 0 mV). EPSCs were likewise isolated by holding cells at IPSC reversal potential (about -75 mV). 2. As stimulus intensities were increased, the magnitude of the postsynaptic currents also increased. Over the range of stimuli applied (2-10 V), EPSCs did not exhibit an upper limit. However, fast gamma-aminobutyric acid-A (GABAA-mediated IPSCs reached a maximum at intensities approximately 2 times threshold. 3. The limit on fast inhibition was unresponsive to alterations in N-methyl-D-aspartate (NMDA)-mediated excitation. Exposure to nominally magnesium-free solutions or to the NMDA antagonist 3-[(RS)-2-carboxypiperazin-4-yl]-propyl-1-phosphonic acid did not affect the fast IPSC maximum. Shifts in the input-output curves for submaximal activation of IPSCs were seen, which were attributed to polysynaptic excitation. 4. Blockade of alpha-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid/kainate (non-NMDA) receptors with 6-cyano-7-nitro-quinoxaline-2,3-dione (CNQX) completely abolished synaptically driven, fast GABAA-mediated inhibition. These findings suggested that neocortical inhibitory cells could be driven exclusively through non-NMDA transmission. 5. By comparison, in hippocampal CA1 pyramidal neurons maximal fast inhibition was sensitive to both NMDA and non-NMDA receptor blockade. 6. The results in neocortex were corroborated by direct intracellular recordings from layer V-VI interneurons. Non-NMDA receptor blockade with CNQX prevented synaptic activation of action potentials in these cells, even during cotreatment with magnesium-free solution. 7. Together, these results suggest that recruitment of GABA(A) ergic IPSCs in neocortex is ultimately driven via glutamatergic afferents arriving at non-NMDA receptors on interneurons. Properties limiting fast inhibition would favor the propagation of enhanced excitatory activity through the neuronal network.

Synaptic response patterns of neurons in the cortex of rat inferior colliculus

Brain functions such as learning and memory rely on synaptic plasticity. Many studies have shown that synaptic plasticity can be driven by the timings between pre- and postsynaptic spikes, also known as spike-timing-dependent plasticity (STDP). In most of the modeling studies exploring STDP functions, presynaptic spikes have been postulated to be Poisson (random) spikes and postsynaptic neurons have been described using an integrate-and-fire model, for simplicity. However, experimental data suggest this is not necessarily true. In this study, we investigated how STDP worked in synaptic competition if more neurophysiologically realistic properties for pre- and postsynaptic dynamics were incorporated; presynaptic (input) spikes obeyed a gamma process and the postsynaptic neuron was a multi-timescale adaptive threshold model. Our results showed that STDP strengthened specific combinations of pre- and postsynaptic properties; regular spiking neurons favored regular input spikes whereas random spiking neurons did random ones, suggesting neural information coding utilizes both the properties.

The influence of gamma-aminobutyric acidB (GABAB) receptor stimulation on the excitatory and inhibitory synaptic potentials and membrane properties of identified striatal spiny neurons was examined in a corticostriatal slice preparation. Stimulation of the subcortical white matter evoked a monosynaptic, excitatory postsynaptic potential (EPSP) and a polysynaptic, inhibitory postsynaptic potential (IPSP) in spiny neurons. The EPSP had two components: a large amplitude response which could be blocked by the kainate/quisqualate receptor antagonist, 6-cyano-7-nitroquinoxaline-2,3-dione (CNQX, 10 microM), and a small amplitude, long-duration depolarization which could be blocked by the N-methyl-D-aspartate receptor antagonist, d-(-)-2-amino-5-phosphonovaleric acid (APV, 100 microM). The IPSP was observed as a membrane depolarization when recorded from neurons at resting membrane potential. However, when neurons were injected with the Na(+)-channel blocker, QX-314, allowing cells to be depolarized above their spike thresholds, a prominent hyperpolarizing IPSP was readily observed which could be blocked by the GABAA antagonist, bicuculline (10-50 microM). This bicuculline-sensitive IPSP was responsible for the inhibition of EPSP amplitude and probability of spike discharge revealed using paired stimulation of the subcortical white matter. The amplitude of both the EPSP and the IPSP were depressed by the GABAB receptor agonist, p-chlophenyl-GABA (baclofen, 0.5-100 microM) in a concentration-dependent manner. Baclofen also blocked paired stimulus inhibition of spike discharge. These effects of baclofen persisted in slices in which the cortex was removed and were reversed by the GABAB receptor antagonist, 3-amino-3-hydroxy-2-(4-chlorophenyl)-propanesulphonic acid (saclofen, 100-500 microM). In contrast to its profound influence on synaptic input, baclofen did not alter resting membrane potential, input resistance, membrane current-voltage relationship, or spike threshold of the cells recorded, and therefore did not appear to exert direct postsynaptic effects on the striatal spiny neurons. Taken together, these data indicate that the depressant effects of baclofen on EPSPs are mediated through GABAB receptors located on the terminals of glutamatergic afferents within the striatum. Moreover, the results suggest that the actions of baclofen on IPSPs and paired stimulus inhibition are produced by activation of GABAB receptors within the striatum at a site presynaptic to spiny neurons, either on the terminals of GABAergic afferents or on an interposed non-spiny GABAergic cell. Thus, GABAB hetero- and auto-receptors have the capacity to provide a negative feedback mechanism through which the major excitatory and inhibitory inputs to striatal spiny neurons are regulated.

1. The interaction between inhibitory interneurons and cortical pyramidal neurons was studied by use of computer simulations to test whether inhibitory interneurons could assist in phase-locking postsynaptic cells. Two models were used: a simplified model, which included only 3 membrane channels, and a detailed 11-channel model. 2. The 11-channel model included most of the ion channels known to be present in neocortical pyramidal neurons as well as calcium diffusion and other membrane mechanisms. The kinetics for the channels were obtained from voltage-clamp studies in a variety of preparations. The parameters were then adjusted to produce repetitive bursting similar to that seen in some cortical pyramidal cells entrained during visual stimulation. 3. Phase-locking to a train of inhibitory postsynaptic potentials (IPSPs) located on or near the soma was observed in the 3-channel model cell subjected to random synaptic bombardment. In the 11-channel model, phase-locking due to multiple IPSPs was compared with phase-locking due to multiple excitatory postsynaptic potentials (EPSPs). Phase-locking began to occur when 20% of the IPSPs (20/100) or 40% of the EPSPs (4,000/10,000) were synchronized. The exact percentages differed with different 11-channel models, but either EPSPs or IPSPs would generally produce entrainment with approximately 40% synchronization. Thus 40 inhibitory boutons had an effect equivalent to 4,000 excitatory boutons in producing phase-locking. 4. Phase-locking with IPSPs in these models was possible because the IPSPs could cause either an increase or a decrease in firing rate over a limited range. The IPSPs served a modulatory role, increasing the rate of firing in some cases and decreasing it in others, depending on the state of the cell. 5. We examined frequency entrainment by IPSPs. In the 3-channel model, frequency entrainment of a postsynaptic cell was observed with a rapid train of strong (20-100 nS), brief, compound IPSPs. A 40-Hz compound IPSP train of 60 nS entrained cells having initial firing rates between 32 and 47 Hz. Below this range, cells could be partially entrained. Above the range, entrainment would fail. Frequency entrainment in the 3-channel model generally occurred on the first cycle after onset of the IPSPs. 6. Phase-locking and frequency entrainment were less robust in the 11-channel model. This was partly because bursts rather than individual spikes were being entrained. A 40-Hz, 90-nS compound IPSP train entrained a model cell upward from 34 Hz. Downward frequency entrainment also occurred.(ABSTRACT TRUNCATED AT 400 WORDS)

Auditory information is carried from the cochlear nuclei to the inferior colliculi through six parallel ascending pathways, one of which is through stellate cells of the ventral cochlear nuclei (VCN) through the trapezoid body. To characterize and identify the synaptic influences on T stellate cells, intracellular recordings were made from anatomically identified stellate cells in parasagittal slices of murine cochlear nuclei. Shocks to the auditory nerve consistently evoked five types of synaptic responses in T stellate cells, which reflect sources intrinsic to the cochlear nuclear complex. 1) Monosynaptic excitatory postsynaptic potentials (EPSPs) that were blocked by 6,7-dinitroquinoxaline-2,3-dione (DNQX), an antagonist of alpha-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors, probably reflected activation by auditory nerve fibers. Electrophysiological estimates indicate that about five auditory nerve fibers converge on one T stellate cell. 2) Disynaptic, glycinergic inhibitory postsynaptic potentials (IPSPs) arise through inhibitory interneurons in the VCN or in the dorsal cochlear nucleus (DCN). 3) Slow depolarizations, the source of which has not been identified, that lasted between 0.2 and 1 s and were blocked by -2-amino-5-phosphonovaleric acid (APV), the N-methyl-D-aspartate (NMDA) receptor antagonist. 4) Rapid, late glutamatergic EPSPs are polysynaptic and may arise from other T stellate cells. 5) Trains of late glycinergic IPSPs after single or repetitive shocks match the responses of D stellate cells, showing that D stellate cells are one source of glycinergic inhibition to T stellate cells. The source of late, polysynaptic EPSPs and IPSPs was assessed electrophysiologically and pharmacologically. Late synaptic responses in T stellate cells were enhanced by repetitive stimulation, indicating that the interneurons from which they arose should fire trains of action potentials in responses to trains of shocks. Late EPSPs and late IPSPs were blocked by APV and enhanced by the removal of Mg2+, indicating that the interneurons were driven at least in part through NMDA receptors. Bicuculline, a gamma-aminobutyric acid-A (GABAA) receptor antagonist, enhanced the late PSPs, indicating that GABAergic inhibition suppresses both the glycinergic interneurons responsible for the trains of IPSPs in T-stellate cells and the interneuron responsible for late EPSPs in T stellate cells. The glycinergic interneurons that mediate the series of IPSPs are intrinsic to the ventral cochlear nucleus because long series of IPSPs were recorded from T stellate cells in slices in which the DCN was removed. These experiments indicate that T stellate cells are a potential source of late EPSPs and that D stellate cells are a potential source for trains of late IPSPs.

Neuromorphic computing has emerged as a highly promising paradigm for energy-efficient and real-time processing of event-driven workloads such as spiking neural networks (SNNs). These systems mimic the sparse and asynchronous communication behavior of the brain, and are particularly well-suited for edge AI, robotics, and low-power sensory processing. However, as neuromorphic processors scale to support larger and more complex SNNs, the communication infrastructure becomes a primary bottleneck. Conventional Network-on-Chip (NoC) architectures--while widely used in many-core systems--are ill-suited for neuromorphic workloads due to their reliance on area and energy intensive buffers, routing tables, and continuous link activity. Moreover, NoCs often fail to accommodate the bursty, multicast, and temporally sensitive nature of SNN traffic, leading to spike loss, latency variation, and unnecessary power dissipation. This dissertation introduces a comprehensive interconnect solution based on the segmented ladder bus, a lightweight and scalable bus architecture optimized for spiking communication. The first contribution is the design of ADIONA, a dynamic segmented ladder bus architecture that leverages bufferless switching and compile-time programmability to enable multiple concurrent paths with low energy and latency. ADIONA's segmented topology and runtime control provide flexibility while maintaining deterministic behavior and hardware simplicity. To support application-level deployment on ADIONA, the second major contribution is MASS--a mapping and scheduling framework for SNNs on segmented buses. MASS includes: (1) a heuristic-based cluster mapping strategy to minimize spike loss and communication energy, (2) a traffic scheduling algorithm to prevent path collisions through temporal grouping, and (3) a path routing method that minimizes path length and congestion. Experimental results show that MASS eliminates spike loss and reduces interconnect energy by up to 13.5% compared to randomized baselines, while supporting high-throughput traffic patterns. The final contribution is an efficient architecture proposal and a hardware realization of the control plane, via a scenario-aware controller implemented on FPGA. To manage control plane scaling complexity, we propose two algorithms--greedy grouping and maximal clique extraction--to compress switch control scenarios and minimize memory usage. Our FPGA results show that the control plane occupies less than 10% of the total area and scales sublinearly with the network size, validating the practicality of the approach. Together, these contributions form a complete and scalable communication framework for neuromorphic systems, spanning architectural design, algorithmic optimization, and hardware implementation. The segmented ladder bus offers a promising path toward future neuromorphic processors that demand low-power, low-latency, and scalable interconnects.

There is a huge need for embedded machine learning for portable devices and smart sensors to power the next generation of Internet of Things (IoT). Implementation of neural networks involve large number of arithmetic and memory operations. Realization of the arithmetic blocks with conventional digital circuits will inherit a trade-off between accuracy of calculation and area required. On the other hand, existing analog building blocks for neural networks suffer from inaccuracies related to process variation and large power consumption. Moreover, the efficiency of computation-memory interface is degrading since memory bandwidth increment is poor compared to computation throughput growth in CMOS technology. Though the recent large scale neuromorphic circuits have used localized random access memory for reducing memory operations, this local memory size will not be scalable with increasing size of datasets. Here, we explore novel CMOS and post-CMOS circuits to realize ultra-low power neuromorphic circuit by co-design of algorithm and hardware. The solutions also overcome the issue of increasing bandwidth gap between memory operations and computation. First, a deep neural network with 2 convolution layers and 2 fully connected layers, is chosen and tuned for hardware implementation. The network has few tunable parameters (_ 40000) and is 40 times faster in training than usual deep neural networks with 4 layers. We propose a compact, single transistor element for realizing the connections inside the neural network- called ’synapse’. These synapses perform the computations involved by virtue of mismatch inherent in their fabrication. We use a current mirror array with n-input lines and m output lines, to perform ’n x m’ Multiply and Accumulation operation. The resultant neuromorphic circuit can emulate multi-layered artificial vision system. A circuit fabricated in 0.35_m CMOS process is characterized and a behavioral model is simulated for the deep neural network. Here, the learning is done offline. The inputs to the network may vary according to environmental conditions, for which we need an adaptive neural network. Hence we propose a second solution where the neuromorphic circuit can adapt it’s parameters in real time. With the advent of novel nanoscale devices with physical properties well matched to neural network enabling computations at energies much lower than CMOS, the research also focuses on use of a post-CMOS spintronic device-a Domain-Wall magnet for obtaining a low power spike timing dependent plastic (STDP) synapse for online learning. The spin-mode signals are injected across small potential (_ 50mV) through multiple layers of ferromagnetic and non-magnetic layers. Here we discuss the implementation of a spiking neural network with synapses which can be trained according to STDP learning rule. A detailed study with the help of device circuit co-simulation is done. Possible use of this synapse in online, real time learning spiking neural networks is also illustrated in this thesis.

We investigated whether entorhinal cortex (EC) layer IV neurons are hyperexcitable in the post-selfsustaining limbic status epilepticus (post-SSLSE) animal model of temporal lobe epilepsy. We studied naive rats (n = 44), epileptic rats that had experienced SSLSE resulting in spontaneous seizures (n = 45), and electrode controls (n = 7). There were no differences between electrode control and naive groups, which were pooled into a single control group. Intracellular and extracellular recordings were made from deep layers of EC, targeting layer IV, which was activated by stimulation of the superficial layers of EC or the angular bundle. There were no differences between epileptic and control neurons in basic cellular characteristics, and all neurons were quiescent under resting conditions. In control tissue, 77% of evoked intracellular responses consisted of a short-duration [8.6 +/- 1.3 (SE) ms] excitatory postsynaptic potential and a single action potential followed by gamma-aminobutyric acid-A (GABAA) and GABAB inhibitory post synaptic potentials (IPSPs). Ten percent of controls did not contain IPSPs. In chronically epileptic tissue, evoked intracellular responses demonstrated prolonged depolarizing potentials (256 +/- 39 ms), multiple action potentials (13 +/- 4), and no IPSPs. Ten percent of epileptic responses were followed by rhythmic "clonic" depolarizations. Epileptic responses exhibited an all-or-none response to progressive increases in stimulus intensity and required less stimulation to elicit action potentials. In both epileptic and control animals, intracellular responses correlated precisely in morphology and duration with extracellular field potentials. Severing the hippocampus from the EC did not alter the responses. Duration of intracellular epileptic responses was reduced 22% by the N-methyl--aspartate (NMDA) antagonist (-)-2-amino-5-phosphonovaleric acid (APV), but they did not return to normal and IPSPs were not restored. Epileptic and control responses were abolished by the non-NMDA antagonist 6, 7-dinitroquinoxaline-2-3-dione (DNQX). A monosynaptic IPSP protocol was used to test connectivity of inhibitory interneurons to primary cells by direct activation of interneurons with a stimulating electrode placed near the recording electrode in the presence of APV and DNQX. Using this protocol, IPSPs similar to control (P > 0.05) were seen in epileptic cells. The findings demonstrate that deep layer EC cells are hyperexcitable or "epileptiform" in this model. Hyperexcitability is not due to interactions with the hippocampus. It is due partially to augmented NMDA-mediated excitation. The lack of IPSPs in epileptic neurons may suggest inhibition is impaired, but we found evidence that inhibitory interneurons are connected to their target cells and are capable of inducing IPSPs.

Spiking neural network (SNN) distinguish themselves from artificial neural network (ANN) because of their inherent temporal processing and spike-based computations, enabling a power-efficient implementation in neuromorphic hardware. In this study, we demonstrate that data processing with spiking neurons can be enhanced by co-learning the synaptic weights with two other biologically inspired neuronal features: (1) a set of parameters describing neuronal adaptation processes and (2) synaptic propagation delays. The former allows a spiking neuron to learn how to specifically react to incoming spikes based on its past. The trained adaptation parameters result in neuronal heterogeneity, which leads to a greater variety in available spike patterns and is also found in the brain. The latter enables to learn to explicitly correlate spike trains that are temporally distanced. Synaptic delays reflect the time an action potential requires to travel from one neuron to another. We show that each of the co-learned features separately leads to an improvement over the baseline SNN and that the combination of both leads to state-of-the-art SNN results on all speech recognition datasets investigated with a simple 2-hidden layer feed-forward network. Our SNN outperforms the benchmark ANN on the neuromorphic datasets (Spiking Heidelberg Digits and Spiking Speech Commands), even with fewer trainable parameters. On the 35-class Google Speech Commands dataset, our SNN also outperforms a GRU of similar size. Our study presents brain-inspired improvements in SNN that enable them to excel over an equivalent ANN of similar size on tasks with rich temporal dynamics.