由于其事件驱动的计算，尖峰神经网络（SNN）已成为常规人工神经网络（ANN）的节能替代方案。考虑到SNN模型的未来部署到限制神经形态设备上，许多研究应用了最初用于ANN模型压缩的技术，例如网络量化，修剪和知识蒸馏，用于SNN。其中，关于知识蒸馏的现有作品报告了学生SNN模型的准确性提高。但是，对能源效率的分析也是SNN的重要特征。在本文中，我们从准确性和能源效率方面彻底分析了蒸馏SNN模型的性能。在此过程中，我们观察到使用常规知识蒸馏方法时，尖峰数量大幅增加，导致能量效率低下。基于此分析，为了达到能源效率，我们提出了一种具有异质温度参数的新知识蒸馏方法。我们在两个不同的数据集上评估我们的方法，并表明由此产生的SNN学生满足了尖峰数量的准确性和减少。在MNIST数据集上，我们提议的学生SNN的精度高达0.09％，与接受常规知识蒸馏方法的学生SNN相比，SNN的峰值降低了65％。我们还将结果与其他SNN压缩技术和训练方法进行了比较。

机器学习中的知识蒸馏是将知识从名为教师的大型模型转移到一个名为“学生”的较小模型的过程。知识蒸馏是将大型网络（教师）压缩到较小网络（学生）的技术之一，该网络可以部署在手机等小型设备中。当教师和学生之间的网络规模差距增加时，学生网络的表现就会下降。为了解决这个问题，在教师模型和名为助教模型的学生模型之间采用了中间模型，这反过来弥补了教师与学生之间的差距。在这项研究中，我们已经表明，使用多个助教模型，可以进一步改进学生模型（较小的模型）。我们使用加权集合学习将这些多个助教模型组合在一起，我们使用了差异评估优化算法来生成权重值。

Despite the fact that deep neural networks are powerful models and achieve appealing results on many tasks, they are too large to be deployed on edge devices like smartphones or embedded sensor nodes. There have been efforts to compress these networks, and a popular method is knowledge distillation, where a large (teacher) pre-trained network is used to train a smaller (student) network. However, in this paper, we show that the student network performance degrades when the gap between student and teacher is large. Given a fixed student network, one cannot employ an arbitrarily large teacher, or in other words, a teacher can effectively transfer its knowledge to students up to a certain size, not smaller. To alleviate this shortcoming, we introduce multi-step knowledge distillation, which employs an intermediate-sized network (teacher assistant) to bridge the gap between the student and the teacher. Moreover, we study the effect of teacher assistant size and extend the framework to multi-step distillation. Theoretical analysis and extensive experiments on CIFAR-10,100 and ImageNet datasets and on CNN and ResNet architectures substantiate the effectiveness of our proposed approach.

One of the most efficient methods for model compression is hint distillation, where the student model is injected with information (hints) from several different layers of the teacher model. Although the selection of hint points can drastically alter the compression performance, conventional distillation approaches overlook this fact and use the same hint points as in the early studies. Therefore, we propose a clustering based hint selection methodology, where the layers of teacher model are clustered with respect to several metrics and the cluster centers are used as the hint points. Our method is applicable for any student network, once it is applied on a chosen teacher network. The proposed approach is validated in CIFAR-100 and ImageNet datasets, using various teacher-student pairs and numerous hint distillation methods. Our results show that hint points selected by our algorithm results in superior compression performance compared to state-of-the-art knowledge distillation algorithms on the same student models and datasets.

Spiking neural networks (SNNs) are receiving increasing attention due to their low power consumption and strong bio-plausibility. Optimization of SNNs is a challenging task. Two main methods, artificial neural network (ANN)-to-SNN conversion and spike-based backpropagation (BP), both have their advantages and limitations. For ANN-to-SNN conversion, it requires a long inference time to approximate the accuracy of ANN, thus diminishing the benefits of SNN. With spike-based BP, training high-precision SNNs typically consumes dozens of times more computational resources and time than their ANN counterparts. In this paper, we propose a novel SNN training approach that combines the benefits of the two methods. We first train a single-step SNN by approximating the neural potential distribution with random noise, then convert the single-step SNN to a multi-step SNN losslessly. The introduction of Gaussian distributed noise leads to a significant gain in accuracy after conversion. The results show that our method considerably reduces the training and inference times of SNNs while maintaining their high accuracy. Compared to the previous two methods, ours can reduce training time by 65%-75% and achieves more than 100 times faster inference speed. We also argue that the neuron model augmented with noise makes it more bio-plausible.

我们提出了一种新的学习算法，使用传统的人工神经网络（ANN）作为代理训练尖刺神经网络（SNN）。我们分别与具有相同网络架构和共享突触权重的集成和火（IF）和Relu神经元进行两次SNN和ANN网络。两个网络的前进通过完全独立。通过假设具有速率编码的神经元作为Relu的近似值，我们将SNN中的SNN的误差进行了回复，以更新共享权重，只需用SNN的ANN最终输出替换ANN最终输出。我们将建议的代理学习应用于深度卷积的SNNS，并在Fahion-Mnist和CiFar10的两个基准数据集上进行评估，分别为94.56％和93.11％的分类准确性。所提出的网络可以优于培训的其他深鼻涕，训练，替代学习，代理梯度学习，或从深处转换。转换的SNNS需要长时间的仿真时间来达到合理的准确性，而我们的代理学习导致高效的SNN，模拟时间较短。

我们如何为神经系统带来隐私和能效？在本文中，我们提出了PrivateNN，旨在从预先训练的ANN模型构建低功耗尖峰神经网络（SNNS），而不会泄漏包含在数据集中的敏感信息。在这里，我们解决两种类型的泄漏问题：1）当网络在Ann-SNN转换过程中访问真实训练数据时，会导致数据泄漏。 2）当类相关的特征可以从网络参数重建时，会导致类泄漏。为了解决数据泄漏问题，我们从预先培训的ANN生成合成图像，并使用所生成的图像将ANN转换为SNNS。然而，转换的SNNS仍然容易受到类泄漏的影响，因为权重参数相对于ANN参数具有相同的（或缩放）值。因此，通过训练SNNS，通过训练基于时间尖峰的学习规则来加密SNN权重。使用时间数据更新权重参数使得SNN难以在空间域中解释。我们观察到，加密的私人没有消除数据和类泄漏问题，略微的性能下降（小于〜2），与标准ANN相比，与标准ANN相比的显着的能效增益（约55倍）。我们对各种数据集进行广泛的实验，包括CiFar10，CiFar100和Tinyimagenet，突出了隐私保留的SNN培训的重要性。

在多种方式知识蒸馏研究的背景下，现有方法主要集中在唯一的学习教师最终产出问题。因此，教师网络与学生网络之间存在深处。有必要强制学生网络来学习教师网络的模态关系信息。为了有效利用从教师转移到学生的知识，采用了一种新的模型关系蒸馏范式，通过建模不同的模态之间的关系信息，即学习教师模级克矩阵。

由于稀疏，异步和二进制事件（或尖峰）驱动加工，尖峰神经网络（SNNS）最近成为深度学习的替代方案，可以在神经形状硬件上产生巨大的能效益。然而，从划痕训练高精度和低潜伏期的SNN，患有尖刺神经元的非微弱性质。要在SNNS中解决此培训问题，我们重新批准批量标准化，并通过时间（BNTT）技术提出时间批量标准化。大多数先前的SNN工程到现在忽略了批量标准化，认为它无效地训练时间SNN。与以前的作品不同，我们提出的BNTT沿着时轴沿着时间轴解耦的参数，以捕获尖峰的时间动态。在BNTT中的时间上不断发展的可学习参数允许神经元通过不同的时间步长来控制其尖峰率，从头开始实现低延迟和低能量训练。我们对CiFar-10，CiFar-100，微小想象特和事件驱动的DVS-CIFAR10数据集进行实验。 BNTT允许我们首次在三个复杂的数据集中培训深度SNN架构，只需25-30步即可。我们还使用BNTT中的参数分布提前退出算法，以降低推断的延迟，进一步提高了能量效率。

由于它们的时间加工能力及其低交换（尺寸，重量和功率）以及神经形态硬件中的节能实现，尖峰神经网络（SNNS）已成为传统人工神经网络（ANN）的有趣替代方案。然而，培训SNNS所涉及的挑战在准确性方面有限制了它们的表现，从而限制了他们的应用。因此，改善更准确的特征提取的学习算法和神经架构是SNN研究中的当前优先级之一。在本文中，我们展示了现代尖峰架构的关键组成部分的研究。我们在从最佳执行网络中凭经验比较了图像分类数据集中的不同技术。我们设计了成功的残余网络（Reset）架构的尖峰版本，并测试了不同的组件和培训策略。我们的结果提供了SNN设计的最新版本，它允许在尝试构建最佳视觉特征提取器时进行明智的选择。最后，我们的网络优于CIFAR-10（94.1％）和CIFAR-100（74.5％）数据集的先前SNN架构，并将现有技术与DVS-CIFAR10（71.3％）相匹配，参数较少而不是先前的状态艺术，无需安静转换。代码在https://github.com/vicenteax/spiking_resnet上获得。

Event-based simulations of Spiking Neural Networks (SNNs) are fast and accurate. However, they are rarely used in the context of event-based gradient descent because their implementations on GPUs are difficult. Discretization with the forward Euler method is instead often used with gradient descent techniques but has the disadvantage of being computationally expensive. Moreover, the lack of precision of discretized simulations can create mismatches between the simulated models and analog neuromorphic hardware. In this work, we propose a new exact error-backpropagation through spikes method for SNNs, extending Fast \& Deep to multiple spikes per neuron. We show that our method can be efficiently implemented on GPUs in a fully event-based manner, making it fast to compute and precise enough for analog neuromorphic hardware. Compared to the original Fast \& Deep and the current state-of-the-art event-based gradient-descent algorithms, we demonstrate increased performance on several benchmark datasets with both feedforward and convolutional SNNs. In particular, we show that multi-spike SNNs can have advantages over single-spike networks in terms of convergence, sparsity, classification latency and sensitivity to the dead neuron problem.

Spiking neural networks (SNN) are a viable alternative to conventional artificial neural networks when energy efficiency and computational complexity are of importance. A major advantage of SNNs is their binary information transfer through spike trains. The training of SNN has, however, been a challenge, since neuron models are non-differentiable and traditional gradient-based backpropagation algorithms cannot be applied directly. Furthermore, spike-timing-dependent plasticity (STDP), albeit being a spike-based learning rule, updates weights locally and does not optimize for the output error of the network. We present desire backpropagation, a method to derive the desired spike activity of neurons from the output error. The loss function can then be evaluated locally for every neuron. Incorporating the desire values into the STDP weight update leads to global error minimization and increasing classification accuracy. At the same time, the neuron dynamics and computational efficiency of STDP are maintained, making it a spike-based supervised learning rule. We trained three-layer networks to classify MNIST and Fashion-MNIST images and reached an accuracy of 98.41% and 87.56%, respectively. Furthermore, we show that desire backpropagation is computationally less complex than backpropagation in traditional neural networks.

较大的尖峰神经网络（SNN）模型通常是有利的，因为它们可以提供更高的精度。但是，在资源和能源约束的嵌入式平台上采用此类模型效率低下。为此，我们提出了一个Tinysnn框架，该框架优化了在训练和推理阶段中SNN处理的记忆和能量需求，同时保持准确性很高。它是通过减少SNN操作，提高学习质量，量化SNN参数并选择适当的SNN模型来实现的。此外，我们的Tinysnn量化了不同的SNN参数（即权重和神经元参数），以最大程度地提高压缩，同时探索量化方案，精度级别和舍入方案的不同组合，以找到提供可接受准确性的模型。实验结果表明，与基线网络相比，我们的Tinysnn显着降低了不准确损失的SNN的记忆足迹和能量消耗。因此，我们的Tinysnn有效地压缩给定的SNN模型，以记忆和节能的方式获得高精度，从而使SNN能够用于资源和能源受限的嵌入式应用程序。

Figure 1. An illustration of standard knowledge distillation. Despite widespread use, an understanding of when the student can learn from the teacher is missing.

Spiking Neural networks (SNN) have emerged as an attractive spatio-temporal computing paradigm for a wide range of low-power vision tasks. However, state-of-the-art (SOTA) SNN models either incur multiple time steps which hinder their deployment in real-time use cases or increase the training complexity significantly. To mitigate this concern, we present a training framework (from scratch) for one-time-step SNNs that uses a novel variant of the recently proposed Hoyer regularizer. We estimate the threshold of each SNN layer as the Hoyer extremum of a clipped version of its activation map, where the clipping threshold is trained using gradient descent with our Hoyer regularizer. This approach not only downscales the value of the trainable threshold, thereby emitting a large number of spikes for weight update with a limited number of iterations (due to only one time step) but also shifts the membrane potential values away from the threshold, thereby mitigating the effect of noise that can degrade the SNN accuracy. Our approach outperforms existing spiking, binary, and adder neural networks in terms of the accuracy-FLOPs trade-off for complex image recognition tasks. Downstream experiments on object detection also demonstrate the efficacy of our approach.

Emergence of deep neural networks (DNNs) has raised enormous attention towards artificial neural networks (ANNs) once again. They have become the state-of-the-art models and have won different machine learning challenges. Although these networks are inspired by the brain, they lack biological plausibility, and they have structural differences compared to the brain. Spiking neural networks (SNNs) have been around for a long time, and they have been investigated to understand the dynamics of the brain. However, their application in real-world and complicated machine learning tasks were limited. Recently, they have shown great potential in solving such tasks. Due to their energy efficiency and temporal dynamics there are many promises in their future development. In this work, we reviewed the structures and performances of SNNs on image classification tasks. The comparisons illustrate that these networks show great capabilities for more complicated problems. Furthermore, the simple learning rules developed for SNNs, such as STDP and R-STDP, can be a potential alternative to replace the backpropagation algorithm used in DNNs.

基于蒸馏的压缩网络的性能受蒸馏质量的管辖。大型网络（教师）到较小网络（学生）的次优蒸馏的原因主要归因于给定教师与学生对的学习能力中的差距。虽然很难蒸馏所有教师的知识，但可以在很大程度上控制蒸馏质量以实现更好的性能。我们的实验表明，蒸馏品质主要受教师响应的质量来限制，这反过来又受到其反应中存在相似信息的影响。训练有素的大容量老师在学习细粒度辨别性质的过程中丢失了类别之间的相似性信息。没有相似性信息导致蒸馏过程从一个例子 - 许多阶级学习减少到一个示例 - 一类学习，从而限制了教师的不同知识的流程。由于隐式假设只能蒸馏出灌输所知，而不是仅关注知识蒸馏过程，我们仔细审查了知识序列过程。我们认为，对于给定的教师 - 学生对，通过在训练老师的同时找到批量大小和时代数量之间的甜蜜点，可以提高蒸馏品。我们讨论了找到这种甜蜜点以便更好地蒸馏的步骤。我们还提出了蒸馏假设，以区分知识蒸馏和正则化效果之间的蒸馏过程的行为。我们在三个不同的数据集中进行我们的所有实验。

脑启发的尖峰神经网络（SNN）已成功应用于许多模式识别域。基于SNN的深层结构在感知任务（例如图像分类，目标检测）中取得了可观的结果。但是，深SNN在加强学习（RL）任务中的应用仍然是一个问题。尽管以前有关于SNN和RL组合的研究，但其中大多数专注于浅网络的机器人控制问题，或使用ANN-SNN转换方法来实施Spiking Spiking Deep Q Network（SDQN）。在这项工作中，我们数学分析了SDQN中尖峰信号特征消失的问题，并提出了一种基于潜在的层归一化（PBLN）方法，以直接训练尖峰尖峰深度Q网络。实验表明，与最先进的ANN-SNN转换方法和其他SDQN作品相比，建议的PBLN Spiking Deep Q Networks（PL-SDQN）在Atari游戏任务上取得了更好的性能。

Spiking Neural Networks (SNNs) are bio-plausible models that hold great potential for realizing energy-efficient implementations of sequential tasks on resource-constrained edge devices. However, commercial edge platforms based on standard GPUs are not optimized to deploy SNNs, resulting in high energy and latency. While analog In-Memory Computing (IMC) platforms can serve as energy-efficient inference engines, they are accursed by the immense energy, latency, and area requirements of high-precision ADCs (HP-ADC), overshadowing the benefits of in-memory computations. We propose a hardware/software co-design methodology to deploy SNNs into an ADC-Less IMC architecture using sense-amplifiers as 1-bit ADCs replacing conventional HP-ADCs and alleviating the above issues. Our proposed framework incurs minimal accuracy degradation by performing hardware-aware training and is able to scale beyond simple image classification tasks to more complex sequential regression tasks. Experiments on complex tasks of optical flow estimation and gesture recognition show that progressively increasing the hardware awareness during SNN training allows the model to adapt and learn the errors due to the non-idealities associated with ADC-Less IMC. Also, the proposed ADC-Less IMC offers significant energy and latency improvements, $2-7\times$ and $8.9-24.6\times$, respectively, depending on the SNN model and the workload, compared to HP-ADC IMC.

尖峰 - 神经网络（SNNS）在边缘设备处具有前景，因为与模拟 - 神经网络（ANN）相比，SNN的事件驱动操作提供了显着较低的功率。虽然很难有效地训练SNN，但是已经开发了许多将培训的ANN转换为SNNS的技术。但是，在转换之后，SNN中的准确性和延迟之间存在权衡关系，在大尺寸数据集中导致诸如想象成的大尺寸数据集之间的相当大。我们提出了一种名为TCL的技术，以缓解权衡问题，使得73.87％（VGG-16）和70.37％（Reset-34）的准确性，在SNNS中的250个周期的中等潜伏期。