A flexible framework for structural plasticity in GPU-accelerated sparse spiking neural networks

Activity-Dependent Dendritic Spine Structural Plasticity Is Regulated by Small GTPase Rap1 and Its Target AF-6

The primary approaches used to train spiking neural networks (SNNs) involve either training artificial neural networks (ANNs) first and then transforming them into SNNs, or directly training SNNs using surrogate gradient techniques. Nevertheless, both of these methods encounter a shared challenge: they rely on frame-based methodologies, where asynchronous events are gathered into synchronous frames for computation. This strays from the authentic asynchronous, event-driven nature of SNNs, resulting in notable performance degradation when deploying the trained models on SNN simulators or hardware chips for real-time asynchronous computation. To eliminate this performance degradation, we propose a hardware-based SNN proxy learning method that is called Chip-In-Loop SNN Proxy Learning (CIL-SPL). This approach effectively eliminates the performance degradation caused by the mismatch between synchronous and asynchronous computations. To demonstrate the effectiveness of our method, we trained models using public datasets such as N-MNIST and tested them on the SNN simulator or hardware chip, comparing our results to those classical training methods.

Kalirin-7 Controls Activity-Dependent Structural and Functional Plasticity of Dendritic Spines

Joint A-SNN: Joint training of artificial and spiking neural networks via self-Distillation and weight factorization

<abstract><p>The training of artificial neural networks (ANNs) with rectified linear unit (ReLU) activation via gradient descent (GD) type optimization schemes is nowadays a common industrially relevant procedure. GD type optimization schemes can be regarded as temporal discretization methods for the gradient flow (GF) differential equations associated to the considered optimization problem and, in view of this, it seems to be a natural direction of research to <italic>first aim to develop a mathematical convergence theory for time-continuous GF differential equations</italic> and, thereafter, to aim to extend such a time-continuous convergence theory to implementable time-discrete GD type optimization methods. In this article we establish two basic results for GF differential equations in the training of fully-connected feedforward ANNs with one hidden layer and ReLU activation. In the first main result of this article we establish in the training of such ANNs under the assumption that the probability distribution of the input data of the considered supervised learning problem is absolutely continuous with a bounded density function that every GF differential equation admits for every initial value a solution which is also unique among a suitable class of solutions. In the second main result of this article we prove in the training of such ANNs under the assumption that the target function and the density function of the probability distribution of the input data are piecewise polynomial that every non-divergent GF trajectory converges with an appropriate rate of convergence to a critical point and that the risk of the non-divergent GF trajectory converges with rate 1 to the risk of the critical point. We establish this result by proving that the considered risk function is <italic>semialgebraic</italic> and, consequently, satisfies the <italic>Kurdyka-Łojasiewicz inequality</italic>, which allows us to show convergence of every non-divergent GF trajectory.</p></abstract>

While machine learning (ML) models are becoming mainstream, including in critical application domains, concerns have been raised about the increasing risk of sensitive data leakage. Various privacy attacks, such as membership inference attacks (MIAs), have been developed to extract data from trained ML models, posing significant risks to data confidentiality. While the predominant work in the ML community considers traditional Artificial Neural Networks (ANNs) as the default neural model, neuromorphic architectures, such as Spiking Neural Networks (SNNs), have recently emerged as an attractive alternative mainly due to their significantly low power consumption. These architectures process information through discrete events, i.e., spikes, to mimic the functioning of biological neurons in the brain. While the privacy issues have been extensively investigated in the context of traditional ANNs, they remain largely unexplored in neuromorphic architectures, and little work has been dedicated to investigating their privacy-preserving properties. In this paper, we investigate the question of whether SNNs have inherent privacy-preserving advantages. Specifically, we investigate SNNs’ privacy properties through the lens of MIAs across diverse datasets, in comparison with ANNs. We explore the impact of different learning algorithms (surrogate gradient and evolutionary learning), programming frameworks (snnTorch, TENNLab, and LAVA), and various parameters on the resilience of SNNs against MIA. Our experiments reveal that SNNs demonstrate consistently superior privacy preservation compared to ANNs, with evolutionary algorithms further enhancing their resilience. For example, on the CIFAR-10 dataset, SNNs achieve an AUC as low as 0.59 compared to 0.82 for ANNs, and on CIFAR-100, SNNs maintain a low AUC of 0.58, whereas ANNs reach 0.88. Furthermore, we investigate the privacy-utility trade-off through Differentially Private Stochastic Gradient Descent (DPSGD), observing that SNNs incur a notably lower accuracy drop than ANNs under equivalent privacy constraints.

Dropout is an Artificial Neural Network (ANN) training technique that has been shown to improve ANN performance across canonical machine learning (ML) datasets. Quantitative Structure Activity Relationship (QSAR) datasets used to relate chemical structure to biological activity in Ligand-Based Computer-Aided Drug Discovery pose unique challenges for ML techniques, such as heavily biased dataset composition, and relatively large number of descriptors relative to the number of actives. To test the hypothesis that dropout also improves QSAR ANNs, we conduct a benchmark on nine large QSAR datasets. Use of dropout improved both enrichment false positive rate and log-scaled area under the receiver-operating characteristic curve (logAUC) by 22-46% over conventional ANN implementations. Optimal dropout rates are found to be a function of the signal-to-noise ratio of the descriptor set, and relatively independent of the dataset. Dropout ANNs with 2D and 3D autocorrelation descriptors outperform conventional ANNs as well as optimized fingerprint similarity search methods.

We present a method for spiking neural network simulation with hardware realistic device non-idealities present in the neuron and synapse circuits as an alternative to time-consumingexpensive spice simulation. Neuromorphic machine learning algorithms for spiking neural networks are often simulatedtested with an ideal mathematical description of spiking neurons and synapses. However, silicon implementations of spiking neurons differ significantly from their ideal mathematical models because of device non-idealities and restrictions in the range of device operation. These non-idealities affect the performance of chip implementations of spiking networks. The dynamical system phase plane of the neuron and synapse circuit is used to create a compact representation of the neuron dynamics that captures device non-idealities for simulation of spiking neural networks. The proposed method would allow hardware-aware optimization of neuromorphic algorithms using standard machine learning tools and provide simulated network prediction close to what would be expected from a chip implementation.

Structural Plasticity describes a form of long-term plasticity, in which the pruning and the creation of synapses lead to the formation of memories in the topology of a network of neurons. In contrast, classical learning rules such as spike-timing dependent plasticity (STDP) focus on changing the efficacy of synapses, for example by looking at the correlation of pre-and post-synaptic activity in spiking neural networks. Typically, prolonged correlated activity leads to a long-term potentiation of the synaptic weight, while anti-correlated activity depresses the weight. We propose a computational model that combines classical learning rules with structural changes in spiking neural network architectures that are based on observations on the morphological changes real biological synapses undergo during their live-cycle. Our model is based on the assumption that newly formed synapses are initially silent, due to their lack of AMPA receptors. In these synapses, only co-activation with other synapses can lead to postsynaptic potentials, and if this co-activation is not present for a critical period, the synapse degenerates again [1]. To study the interaction of structural plasticity and classical STDP learning rules, we simulated a highly recurrently connected spiking neural network and presented topological inputs to its neurons. We implemented the triplet STDP learning rule proposed by Pfister and Gerstner [2], and applied a structural plasticity rule where a critical period is opened whenever a synaptic weight is decreased below a certain threshold. If the weight does not manage to reach a set threshold by the end of the critical period, the synapse is pruned, and a new synapse is instantiated within the network; otherwise the synapse is maintained. This approach implies a homeostasis in the number of consolidated synapses in the network, while keeping the connectivity at a desired level of sparseness. We show in Figure ​Figure11 simulation results in which the input topology of the network is first learned using only STDP, and then, after activating structural plasticity, the structure of the connectivity matrix itself is adapted such that it reflects the input topology. Figure 1 Connectivity matrix from input to target neuron population with and without structural plasticity. Top: Randomly initialized connectivity, learned using triplet STDP. Bottom: Connectivity learned using structural plasticity and triplet-STDP. A major advantage of structural plasticity in artificial neural networks is given by the fact that it allows a drastic increase in performance given a finite number of synaptic resources. In addition to offering a promising approach for optimizing performance in software simulated networks, the model we propose optimizes the usage of resources in dedicated hardware neural network implementations that are faced with limited resources for emulating or simulating synaptic connections. This is particularly relevant for electronic implementations of spiking neural networks, ranging from GPU-based systems to mixed signal analog-digital neuromorphic VLSI devices.

In recent years, the scientific community's interest for developing software or hardware tools capable of processing and displaying real-time simulations of large-scale Spiking Neural Networks (SNN) has been growing exponentially. Visualization of the simulations is crucial to observe and study the dynamics of the SNN, because these models involve high level of biological realism by encoding spatial-temporal information into action potentials or spikes like biological neurons do. Existing tools are based on supercomputers or customized neuromorphic architectures to compute large-scale SNN models. However, many of them are unsuitable to carry out real-time analysis of large scale SNN with minimal cost as large amount of complex data needs to be processed and displayed at the same time. This work intends to provide an efficient cost effective and power efficient emulation tool based on hardware-software hybrid to process, control, and display large-scale SNN network behavior in real time. The solution consists of a customized hardware architecture that computes the neural-synaptic parameters and a software environment that displays the same on the screen of a general purpose computer (Host). The data to be visualized is transferred to the host through high speed Ethernet link. The proposed solution tailors the use of the Ethernet link bandwidth to an optimum such that the customized neuromorphic architecture and the host computer duo can be used to visualize the dynamics of SNN network of any size in real-time. The hardware and software platforms are scalable in terms of number of neurons and synapses. This solution gives flexibility to the user to visualize and analyze SNN network of any size avoiding saturation of the high-speed serial Ethernet link while supporting the visualization of SNN dynamics with a good screen resolution.

Spike-based brain-inspired systems have shown an immense capability to achieve internal stability, widely referred to as homeostasis. This ability enrols them as the best candidate for next-generation computational neuroscience as they bridge the gap between neuroscience and machine learning. Spiking Neural Networks (SNN), a third generation Artificial Neural Network (ANN), which operates using discrete events of spikes, contributes to a category of biologically-realistic models of neurons to carry out computations. Spiking Astrocyte-Neuron Networks (SANN) have a characteristic attribute homologous to brain self-repair. Although SNNs are more powerful in theory than 2nd generation ANNs, they are not widely in use as their implementations on normal hardware are computationally-intensive. On the contrary, due to the capability of modern hardware such as FPGAs, which operates in MHz and GHz range, facilitates real-time and faster-than-real-time simulations of SNNs. In this work, we overcome the computational overhead of the SNNs using the benefits of real-time hardware computations, utilizing time-multiplexing to design a Self-rePairing spiking Astrocyte Neural NEtwoRk (SPANNER) chip, generic to users‘ choice of task, emphasizing fault-tolerance, targeting safety-critical applications. We demonstrate the proposed methodology on a SANN system implemented on Xilinx Artix-7 FPGA. The proposed architecture has minimal hardware footprints, power dissipation profile and real-time computational capability, enhancing its usability in constrained applications.

Experience alters brain structure, but the underlying mechanism remained unknown. Structural plasticity reveals that brain function is encoded in generative changes to cells that compete with destructive processes driving neurodegeneration. At an adult critical period, experience increases fiber number and brain size in Drosophila. Here, we asked if Toll receptors are involved. Tolls demarcate a map of brain anatomical domains. Focusing on Toll-2, loss of function caused apoptosis, neurite atrophy and impaired behaviour. Toll-2 gain of function and neuronal activity at the critical period increased cell number. Toll-2 induced cycling of adult progenitor cells via a novel pathway, that antagonized MyD88-dependent quiescence, and engaged Weckle and Yorkie downstream. Constant knock-down of multiple Tolls synergistically reduced brain size. Conditional over-expression of Toll-2 and wek at the adult critical period increased brain size. Through their topographic distribution, Toll receptors regulate neuronal number and brain size, modulating structural plasticity in the adult brain.

The biological brain has self-adaptive ability through neural information processing and regulation. Drawing from the advantage of the biological brain, it is significant to research the robustness of artificial neural network (ANN) based on brain-like intelligence. In this study, based on the Izhikevich neuron model and the synaptic plasticity model which contains excitatory and inhibitory synapses, a spiking neural network (SNN) with small-world topology and a SNN with scale-free topology are constructed. The anti-injury function of two complex SNNs (CSNN) under random attack is comparatively analyzed. On this basis, the information processing of CSNN under attack is further discussed, and the anti-injury mechanism of CSNN is explored based on the synaptic plasticity. The experimental results show that: (1) scale-free SNN (SFSNN) has better performance than small-world SNN (SWSNN) in the anti-injury ability under random attacks. (2) The information processing of CSNN under random attacks is clarified by the linkage effects of dynamic changes in neuron firing, synaptic weight, and topological characteristics. (3) The anti-injury ability of CSNNs is closely related to the dynamic evolution of synaptic weight, which implies the dynamic regulation of synaptic plasticity is the intrinsic factor of the anti-injury function of CSNNs. This study lays a theoretical foundation for the application of brain-like intelligence with adaptive fault-tolerance.

Pursuing intelligence is a long-term goal of the human, towards which two routes have been paved on the road: neuromorphic computing driven by neuroscience and machine learning driven by computer science (Pei et al., 2019). Spiking neural networks (SNNs) and neuromorphic chips (Basu et al., 2022;Christensen et al., 2022) dominate the neuromorphic computing domain, while artificial neural networks (ANNs) and machine learning accelerators (Deng et al., 2020) dominate the machine learning domain. Neuromorphic computing with efficient models and hardware has shown energy efficiency superiority (Renner et al., 2021), however, still lies in its infant stage and presents a gap in terms of accuracy and applications compared to the mature machine learning ecosystem.To this end, we proposed a Research Topic, named "Understanding and Bridging the Gap between Neuromorphic Computing and Machine Learning", in Frontiers in Neuroscience and Frontiers in Computational Neuroscience in 2019, and have successfully published 14 papers on neuromorphic computing and machine learning (Deng et al., 2021). Encouraged by such positive impetus for the neuromorphic computing community, we relaunched the Research Topic in 2022.This time, we have accepted 11 submissions in the end. The scope of these works covers neuromorphic models and algorithms, hardware implementation, and programming frameworks.SNNs encode information in spike events and process information using neural dynamics, which differ from ANNs. Due to the complicated spatiotemporal dynamics and non-differentiable spike activities, the SNN domain uses plasticitybased unsupervised learning algorithms (Diehl and Cook, 2015) for a long time but suffers from low accuracy. To break the bottleneck of lacking effective learning algorithms, the sophisticated backpropagation method in machine learning has Neuromorphic models enjoy low computational costs owing to the binary spike representation and sparse operations. However, directly executing SNNs on GPUs without tailored optimization is inefficient. Neuromorphic hardware is designed for the efficient execution of SNNs via event-driven computing (Merolla et al., 2014). The decentralized manycore architecture with high computing parallelism and memory locality is widely adopted by neuromorphic chips. However, its fragmented memories and decentralized execution lowers the resource utilization and processing efficiency. Wang et al., 2023 propose the mapping limit concept which points out the resource saving upper limit during logical and physical mapping when deploying neural networks onto neuromorphic chips. A closed-loop mapping strategy with an asynchronous 4D model partition for logical mapping and a Hamilton loop algorithm (HLA) for physical mapping are elaborated. Their methods and performance gains are validated on the TianjicX neuromorphic chip (Ma et al., 2022), which is helpful for building a general and efficient mapping framework for neuromorphic hardware.Software is one of the key components in the ecosystem of neuromorphic computing (Fang et al., 2023), which is sometimes more important than the hardware itself because it determines how much practical efficiency we can gain from the peak efficiency of hardware. In this Research Topic, we accepted one paper on the programming framework for neuromorphic models in this Research Neuromorphic computing is a neuroscience-driven domain in pursuing brain-like intelligence, which is an important route distinct from machine learning. Although neuromorphic systems have not yet demonstrated superior performance over machine learning systems in main stream intelligent tasks, we believe it can be significantly improved when the neuromorphic ecosystem is constructed and becomes iterative between algorithms, models, hardware, software, and benchmarks. This Research Topic is a quite minor step. We hope future works can really bridge the gap between neuromorphic computing and machine learning, along the way to reach the long-term goal of mimicking brain intelligence.

The authors have developed an automated computeraided diagnostic (CAD) scheme by using artificial neural networks (ANNs) on quantitative analysis of image data. Three separate ANNs were applied for detection of interstitial disease on digitized chest images. The first ANN was trained with horizontal profiles in regions of interest (ROIs) selected from normal and abnormal chest radiographs for distinguishing between normal and abnormal patterns. For training and testing of the second ANN, the vertical output patterns obtained from the 1st ANN were used for each ROI. The output value of the second ANN was used to distinguish between normal and abnormal ROIs with interstitial infiltrates. If the ratio of the number of abnormal ROIs to the total number of all ROIs in a chest image was greater than a specified threshold level, the image was classified as abnormal. In addition, the third ANN was applied to distinguish between normal and abnormal chest images. The combination of the rule-based method and the third ANN also was applied to the classification between normal and abnormal chest images. The performance of the ANNs was evaluated by means of receiver operating characteristic (ROC) analysis. The average Az value (area under the ROC curve) for distinguishing between normal and abnormal cases was 0.976 +/- 0.012 for 100 chest radiographs that were not used in training of ANNs. The results indicate that the ANN trained with image data can learn some statistical properties associated with interstitial infiltrates in chest radiographs.