Memory-Dependent Computation and Learning in Spiking Neural Networks Through Hebbian Plasticity.

Spiking Neural Networks (SNNs), widely known as the third generation of neural networks, encode input information temporally using sparse spiking events, which can be harnessed to achieve higher computational efficiency for cognitive tasks. However, considering the rapid strides in accuracy enabled by state-of-the-art Analog Neural Networks (ANNs), SNN training algorithms are much less mature, leading to accuracy gap between SNNs and ANNs. In this paper, we propose different SNN training methodologies, varying in degrees of biofidelity, and evaluate their efficacy on complex image recognition datasets. First, we present biologically plausible Spike Timing Dependent Plasticity (STDP) based deterministic and stochastic algorithms for unsupervised representation learning in SNNs. Our analysis on the CIFAR-10 dataset indicates that STDP-based learning rules enable the convolutional layers to self-learn low-level input features using fewer training examples. However, STDP-based learning is limited in applicability to shallow SNNs (≤4 layers) while yielding considerably lower than state-of-the-art accuracy. In order to scale the SNNs deeper and improve the accuracy further, we propose conversion methodology to map off-the-shelf trained ANN to SNN for energy-efficient inference. We demonstrate 69.96% accuracy for VGG16-SNN on ImageNet. However, ANN-to-SNN conversion leads to high inference latency for achieving the best accuracy. In order to minimize the inference latency, we propose spike-based error backpropagation algorithm using differentiable approximation for the spiking neuron. Our preliminary experiments on CIFAR-10 show that spike-based error backpropagation effectively captures temporal statistics to reduce the inference latency by up to 8× compared to converted SNNs while yielding comparable accuracy

Ca(2+) influx through postsynaptic Ca(v)1.x L-type voltage-gated channels (LTCCs) is particularly effective in activating neuronal biochemical signaling pathways that might be involved in Hebbian synaptic plasticity (i.e., long-term potentiation and depression) and learning and memory. Here, we demonstrate that Ca(v)1.2 is the functionally relevant LTCC isoform in the thalamus-amygdala pathway of mice. We further show that acute pharmacological block of LTCCs abolishes Hebbian plasticity in the thalamus-amygdala pathway and impairs the acquisition of conditioned fear. On the other hand, chronic genetic loss of Ca(v)1.2 triggers a homeostatic change of the synapse, leading to a fundamental alteration of the mechanism of Hebbian plasticity by synaptic incorporation of Ca(2+)-permeable, GluA2-lacking AMPA receptors. Our results demonstrate for the first time the importance of the Ca(v)1.2 LTCC subtype in synaptic plasticity and fear memory acquisition.

Developmental plasticity plays a prominent role in shaping the brain's structure during ongoing learning in response to dynamically changing environments. However, the existing network compression methods for deep artificial neural networks (ANNs) and spiking neural networks (SNNs) draw little inspiration from brain's developmental plasticity mechanisms, thus limiting their ability to learn efficiently, rapidly, and accurately. This paper proposed a developmental plasticity-inspired adaptive pruning (DPAP) method, with inspiration from the adaptive developmental pruning of dendritic spines, synapses, and neurons according to the "use it or lose it, gradually decay" principle. The proposed DPAP model considers multiple biologically realistic mechanisms (such as dendritic spine dynamic plasticity, activity-dependent neural spiking trace, and local synaptic plasticity), with additional adaptive pruning strategy, so that the network structure can be dynamically optimized during learning without any pre-training and retraining. Extensive comparative experiments show consistent and remarkable performance and speed boost with the extremely compressed networks on a diverse set of benchmark tasks for deep ANNs and SNNs, especially the spatio-temporal joint pruning of SNNs in neuromorphic datasets. This work explores how developmental plasticity enables complex deep networks to gradually evolve into brain-like efficient and compact structures, eventually achieving state-of-the-art (SOTA) performance for biologically realistic SNNs.

As computational models inspired by the biological neural system, spiking neural networks (SNN) continue to demonstrate great potential in the landscape of artificial intelligence, particularly in tasks such as recognition, inference, and learning. While SNN focuses on achieving high-level intelligence of individual creatures, Swarm Intelligence (SI) is another type of bio-inspired models that mimic the collective intelligence of biological swarms, i.e., bird flocks, fish school and ant colonies. SI algorithms provide efficient and practical solutions to many difficult optimization problems through multi-agent metaheuristic search. Bridging these two distinct subfields of artificial intelligence has the potential to harness collective behavior and learning ability of biological systems. In this work, we explore the feasibility of connecting these two models by implementing a generalized SI model on SNN. In the proposed computing paradigm, we use SNNs to represent agents in the swarm and encode problem solutions with the spike firing rate and with spike timing. The coupled neurons communicate and modulate each other's action potentials through event-driven spikes and synchronize their dynamics around the states of optimal solutions. We demonstrate that such an SI-SNN model is capable of efficiently solving optimization problems, such as parameter optimization of continuous functions and a ubiquitous combinatorial optimization problem, namely, the traveling salesman problem with near-optimal solutions. Furthermore, we demonstrate an efficient implementation of such neural dynamics on an emerging hardware platform, namely ferroelectric field-effect transistor (FeFET) based spiking neurons. Such an emerging in-silico neuron is composed of a compact 1T-1FeFET structure with both excitatory and inhibitory inputs. We show that the designed neuromorphic system can serve as an optimization solver with high-performance and high energy-efficiency.

The investigation of nonlinear processes in brain systems based on the oscillatory-wave approach is currently one of the most consequential domains of researching signal generation and processing mechanisms in the brain. These investigations may offer insights into various phenomena, such as the mechanism of associative memory, based on network models of biologically relevant neurons. A neural network architecture is proposed in this paper to solve applied problems, such as signal filtering, processing, and recognition of information images. Mechanisms for incomplete fragments were implemented, enabling the extraction and complete restoration of objects from memory. Each neuron in this neural network comprises a Hodgkin–Huxley biophysical model with a Mainen modification, the dynamics of which most plausibly reproduce the processes in the neural cells of the brain. In the research on individual neurons, a constant external current with varying amplitudes was applied. The study determined the average periods of self-oscillations and identified parameter values for the Andronov–Hopf bifurcation. In particular, a stable limit cycle is created and destroyed depending on different scenarios when the bias current is increased or decreased. A stable limit cycle arises via the Andronov–Hopf bifurcation and is terminated via the saddle-node bifurcation on the cycle. The study identified parameter regions whereby two neurons with excitatory and inhibitory synaptic connections synchronize. The neurons were all in self-oscillating mode and exhibited a stable limit cycle in the phase space. A three-layer neural network was constructed, comprising of a reference neuron, sensory, control, and interneurons. The control and intermediate layer connections are arranged based on the Hebbian learning rule. Areas of excitation for various types of neurons were found. Experiments were conducted using the neural network architecture to distinguish binary information patterns encoded through signal phase. Excitatory and inhibitory synaptic connections were employed to encode the images stored in memory. The dynamic patterns are determined through the in-phase or anti-phase mode of phase locking, i.e., phase synchronization relative to the global rhythm. Before reaching the output layer of neurons, the signals are summed on the interneuron layer with the signal of the reference neuron, resulting in the filtering out of a certain information segment (in-phase or anti-phase). The network accurately recognized the patterns after parameter adjustments were made. Areas of neuron phase synchronization were discovered, enabling control of network activity modes. The confirmation of the viability of phase capture and pattern recognition modes using Hodgkin–Huxley–Mainen neurons was achieved. In total, as a result of the implementation of the spike neural network modeling project: 1) two-parameter diagrams of regions of neuronal excitation in various modes are calculated; 2) the effect of excitatory and inhibitory coupling in synaptic currents is used to represent the input signal; 3) an abstract mathematical algorithm for calculating the Hebbian matrix has been transferred to the implementation of synaptic currents between layers of neurons; 4) the effect of phase clusters is used to represent the output pattern; 5) the neurons’ topology in the visual-cerebral department served as a working model for recognizing graphic images through the Hopfield network on a spike neural network, using the Hebbian rule.

Growing experimental evidence shows that both homeostatic and Hebbian synaptic plasticity can be expressed presynaptically as well as postsynaptically. In this review, we start by discussing this evidence and methods used to determine expression loci. Next, we discuss the functional consequences of this diversity in pre- and postsynaptic expression of both homeostatic and Hebbian synaptic plasticity. In particular, we explore the functional consequences of a biologically tuned model of pre- and postsynaptically expressed spike-timing-dependent plasticity complemented with postsynaptic homeostatic control. The pre- and postsynaptic expression in this model predicts (i) more reliable receptive fields and sensory perception, (ii) rapid recovery of forgotten information (memory savings), and (iii) reduced response latencies, compared with a model with postsynaptic expression only. Finally, we discuss open questions that will require a considerable research effort to better elucidate how the specific locus of expression of homeostatic and Hebbian plasticity alters synaptic and network computations.This article is part of the themed issue ‘Integrating Hebbian and homeostatic plasticity’.

Flow of energy and free energy minimization underpins almost every aspect of naturally occurring physical mechanisms. Inspired by this fact this work establishes an energy-based framework that spans the multi-scale range of biological neural systems and integrates synaptic dynamic, synchronous spiking activity and neural states into one consistent working paradigm. Following a bottom-up approach, a hypothetical energy function is proposed for dynamic synaptic models based on the theoretical thermodynamic principles and the Hopfield networks. We show that a synapse exposes stable operating points in terms of its excitatory postsynaptic potential as a function of its synaptic strength. We postulate that synapses in a network operating at these stable points can drive this network to an internal state of synchronous firing. The presented analysis is related to the widely investigated temporal coherent activities (cell assemblies) over a certain range of time scales (binding-by-synchrony). This introduces a novel explanation of the observed (poly)synchronous activities within networks regarding the synaptic (coupling) functionality. On a network level the transitions from one firing scheme to the other express discrete sets of neural states. The neural states exist as long as the network sustains the internal synaptic energy.

With a similar mechanism of processing spikes as biological neural systems, spiking neural networks (SNNs) are expected to perform a better performance than traditional rate based neural networks. Despite the progress in the application of SNNs, how does the brain learn information from spiking based codes remains an open question. To train a spiking neuron output desired spike trains, many learning methods have been proposed, and can be mainly divided into two classes: spike-driven methods and membrane potential-driven methods. Moreover, different learning algorithms have different advantages in different applications. Therefore, in this paper, we analysis and compare the existing learning algorithm for spiking neurons, including learning efficiency, learning accuracy.

Neuromorphic systems (hardware neural networks) derive inspiration from biological neural systems and are expected to be a computing breakthrough beyond conventional von Neumann architecture. Interestingly, in neuromorphic systems, the processing and storing of information can be performed simultaneously by modulating the connection strength of a synaptic device (i.e., synaptic weight). Previously investigated synaptic devices can emulate the functionality of biological synapses successfully by utilizing various nano-electronic phenomena; however, the impact of intrinsic synaptic device variability on the system performance has not yet been studied. Here, we perform a device-to-system level simulation of different synaptic device variation parameters in a designed neuromorphic system that has the potential for unsupervised learning and pattern recognition. The effects of variations in parameters such as the weight modulation nonlinearity (NL), the minimum-maximum weight (Gmin and Gmax), and the weight update margin (ΔG) on the pattern recognition accuracy are analyzed quantitatively. These simulation results can provide guidelines for the continued design and optimization of a synaptic device for realizing a functional large-scale neuromorphic computing system.

Monitoring heart conditions through electrocardiography (ECG) has been the cornerstone of identifying cardiac irregularities. Cardiologists often rely on a detailed analysis of ECG recordings to pinpoint deviations that are indicative of heart anomalies. This traditional method, while effective, demands significant expertise and is susceptible to inaccuracies due to its manual nature. In the realm of computational analysis, Artificial Neural Networks (ANNs) have gained prominence across various domains, which can be attributed to their superior analytical capabilities. Conversely, Spiking Neural Networks (SNNs), which mimic the neural activity of the brain more closely through impulse-based processing, have not seen widespread adoption. The challenge lies primarily in the complexity of their training methodologies. Despite this, SNNs offer a promising avenue for energy-efficient computational models capable of displaying a high-level performance. This paper introduces an innovative approach employing SNNs augmented with an attention mechanism to enhance feature recognition in ECG signals. By leveraging the inherent efficiency of SNNs, coupled with the precision of attention modules, this model aims to refine the analysis of cardiac signals. The novel aspect of our methodology involves adapting the learned parameters from ANNs to SNNs using leaky integrate-and-fire (LIF) neurons. This transfer learning strategy not only capitalizes on the strengths of both neural network models but also addresses the training challenges associated with SNNs. The proposed method is evaluated through extensive experiments on two publicly available benchmark ECG datasets. The results show that our model achieves an overall accuracy of 93.8% on the MIT-BIH Arrhythmia dataset and 85.8% on the 2017 PhysioNet Challenge dataset. This advancement underscores the potential of SNNs in the field of medical diagnostics, offering a path towards more accurate, efficient, and less resource-intensive analyses of heart diseases.

Learning or the formation of memories is a key component of biological intelligence. One distinguishes between different types of memory as, for instance, the declarative memory denoting the storage of facts and concepts which are extracted from perceived stimuli. Thereby, an important question is how the brain distinguishes between stimuli that should be associated and those to be discriminated. Investigating this mechanism, in psychophysical experiments, rats were exposed to gradually morphed environments while recording activity in hippocampal place cells [1,2]. Stability of the recorded place-dependent activation patterns was dependent on the order in which the different environments were explored. In other words, stimulus discrimination or association depends on the order of presentation. Theoretical models of memory rely on the hypothesis that memorization is implemented through activity-dependent synaptic plasticity. Environmental patterns are reflected by correlated activity in subgroups of neurons leading to high recurrent synaptic efficacies among them. These groups of highly interconnected neurons are named cell assemblies (CAs). This long-standing concept with its basic ability of pattern completion is a promising approach to explain several cognitive phenomena [3]. Combining the theory of CAs with the principle of self-organizing maps (SOMs) [4], we present a biologically plausible network model able to reproduce the upper experimental findings. Basically our model consists of three populations of neurons. An input population projects through random excitatory connections on a recurrent layer. Neurons within the layer interact by excitatory synapses. Competition is introduced by an inhibitory population mutually interconnected with every layer neuron. For all neurons, we use a rate-coded neuron model with sigmoidal transfer functions. The plasticity model combines (Hebbian) synaptic plasticity and synaptic scaling (SPASS) [5]. Stimuli are inserted as activity patterns in the input population passing them on via afferent plastic synapses. Recurrent lateral connections in the layer reach a circular neighborhood of fixed range and are also plastic. The incoming excitatory synapses the inhibitory population receives as well as its outgoing connections are non-plastic. In agreement with SOM, our concept maps sufficiently differing stimuli to different representations (CAs) in the layer. Comparable to the experimental procedures described above, we define two non-overlapping activity patterns A and B. According with [1], first we transform stimulus A gradually into B with a short pause between each step of the transition sequence. After A lead to the formation of one memory representation, or rather a single CA, this CA is maintained throughout the following learning protocol. Secondly, the protocol of [2] demands a random presentation of the transition patterns. In this case, we observe the formation of multiple representations. Thus, our model reproduces the experimental findings described above. In contrast to existing theoretical studies [6], it does not require stimulus-dependent modifications of the plasticity rule itself. It further allows for investigating the influence of different variations in the learning protocol as, for instance, of the stimulus presentation time.

A problem with Hebb and local spikes

Machine learning can be effectively applied in control loops to make optimal control decisions robustly. There is increasing interest in using spiking neural networks (SNNs) as the apparatus for machine learning in control engineering because SNNs can potentially offer high energy efficiency, and new SNN-enabling neuromorphic hardware is being rapidly developed. A defining characteristic of control problems is that environmental reactions and delayed rewards must be considered. Although reinforcement learning (RL) provides the fundamental mechanisms to address such problems, implementing these mechanisms in SNN learning has been underexplored. Previously, spike-timing-dependent plasticity learning schemes (STDP) modulated by factors of temporal difference (TD-STDP) or reward (R-STDP) have been proposed for RL with SNN. Here, we designed and implemented an SNN controller to explore and compare these two schemes by considering cart-pole balancing as a representative example. Although the TD-based learning rules are very general, the resulting model exhibits rather slow convergence, producing noisy and imperfect results even after prolonged training. We show that by integrating the understanding of the dynamics of the environment into the reward function of R-STDP, a robust SNN-based controller can be learned much more efficiently than TD-STDP.

Spiking Neural Networks (SNNs) bear the potential of efficient unsupervised\nand continual learning capabilities because of their biological plausibility,\nbut their complexity still poses a serious research challenge to enable their\nenergy-efficient design for resource-constrained scenarios (like embedded\nsystems, IoT-Edge, etc.). We propose SpikeDyn, a comprehensive framework for\nenergy-efficient SNNs with continual and unsupervised learning capabilities in\ndynamic environments, for both the training and inference phases. It is\nachieved through the following multiple diverse mechanisms: 1) reduction of\nneuronal operations, by replacing the inhibitory neurons with direct lateral\ninhibitions; 2) a memory- and energy-constrained SNN model search algorithm\nthat employs analytical models to estimate the memory footprint and energy\nconsumption of different candidate SNN models and selects a Pareto-optimal SNN\nmodel; and 3) a lightweight continual and unsupervised learning algorithm that\nemploys adaptive learning rates, adaptive membrane threshold potential, weight\ndecay, and reduction of spurious updates. Our experimental results show that,\nfor a network with 400 excitatory neurons, our SpikeDyn reduces the energy\nconsumption on average by 51% for training and by 37% for inference, as\ncompared to the state-of-the-art. Due to the improved learning algorithm,\nSpikeDyn provides on avg. 21% accuracy improvement over the state-of-the-art,\nfor classifying the most recently learned task, and by 8% on average for the\npreviously learned tasks.\n

In an artificial Spiking Neural Network (SNN) the information processing and transmission are carried out by spike trains in a manner similar to the generic biological neurons. Recently it has been reported that they are computationally more powerful than the conventional neural networks. Yet, there are no well defined efficient methods for learning due to their rather intricately discontinuous and nonlinear mechanisms. In this paper, we consider a recurrent SNN constructed with integrate-and-fire type spiking neurons. First we propose a learning method such that the SNN possesses desired transient responses (spike-train outputs) by changing the synaptic weights. Further by including periodic state conditions we propose a learning method such that the SNN possesses desired oscillatory responses (limit cycle spike train) by changing both the synaptic weights and the initial conditions. Simulation examples are also provided to verify the efficiency and the applicability of the proposed algorithm.