Hardware Implementation of an Approximate Simplified Piecewise Linear Spiking Neuron

Neuromorphic computing is a promising technology for future machine learning research, which is envisioned to provide lower power consumption, fault tolerance, and massive parallelism. However, the neuromorphic network, i.e., spike neural network (SNN), suffers from a broader range of its application compared with artificial neural networks (ANN). In this work, we propose a stochastic computation SC-based hybrid ANN and SNN scheme, which supports both ANN and SNN computation in SC domain. The ANN computation is converted to SC, which is implemented with simple logic gates. Moreover, we propose a new two-stage stochastic computation that is compatible with the sequence computation style of SNN. Consequently, the logic resource for the hybrid SNN and ANN can be shared. Compared to the state-of-the-art hybrid scheme, the proposed SC-based ANN and SNN exhibit much lower hardware costs and higher hardware efficiency.

S[formula omitted]NN: Time step reduction of spiking surrogate gradients for training energy efficient single-step spiking neural networks

Leaky integrate-and-fire neurons based on perovskite memristor for spiking neural networks

Building accurate and efficient deep neural network (DNN) models for intelligent sensing systems to process data locally is essential. Spiking neural networks (SNNs) have gained significant popularity in recent years because they are more biological-plausible and energy-efficient than DNNs. However, SNNs usually have lower accuracy than DNNs. In this paper, we propose to use SNNs for image sensing applications. Moreover, we introduce the DNN-SNN knowledge distillation algorithm to reduce the accuracy gap between DNNs and SNNs. Our DNNSNN knowledge distillation improves the accuracy of an SNN by transferring knowledge between a DNN and an SNN. To better transfer the knowledge, our algorithm creates two learning paths from a DNN to an SNN. One path is between the output layer and another path is between the intermediate layer. DNNs use real numbers to propagate information between neurons while SNNs use 1-bit spikes. To empower the communication between DNNs and SNNs, we utilize a decoder to decode spikes into real numbers. Also, our algorithm creates a learning path from an SNN to a DNN. This learning path better adapts the DNN to the SNN by allowing the DNN to learn the knowledge from the SNN. Our SNN models are deployed on Loihi, which is a specialized chip for SNN models. On the MNIST dataset, our SNN models trained by the DNN-SNN knowledge distillation achieve better accuracy than the SNN models on GPU trained by other training algorithms with much lower energy consumption per image.

Spiking neural networks (SNN) are a promising approach for low-power edge AI (artificial intelligence), especially when run on dedicated neuromorphic hardware. In this work we set up a SNN in TensorFlow, directly train it in a supervised manner with backpropagation through time (BPTT) and surrogate gradients, and compare it with traditional neural networks, based on convolutional neural networks (CNN) and long-short term memory (LSTM) cells, for radar-based hand-gesture recognition. We demonstrate that a small SNN with only 30 hidden leaky integrate-and-fire (LIF) neurons and threshold encoding can achieve an accuracy of 98.1%. With the more complex adaptive LIF neuron, the activity can be reduced by up to 37.9% without significant loss in accuracy. A comparison to traditional LSTM-networks shows the superiority of the SNN in terms of accuracy and computation cost, indicating that they are a considerable alternative to LSTM-based approaches.

Spiking neural networks (SNNs) have attracted many researchers’ interests due to its biological plausibility and event-driven characteristic. In particular, recently, many studies on high-performance SNNs comparable to the conventional analog-valued neural networks (ANNs) have been reported by converting weights trained from ANNs into SNNs. However, unlike ANNs, SNNs have an inherent latency that is required to reach the best performance because of differences in operations of neuron. In SNNs, not only spatial integration but also temporal integration exists, and the information is encoded by spike trains rather than values in ANNs. Therefore, it takes time to achieve a steady-state of the performance in SNNs. The latency is worse in deep networks and required to be reduced for the practical applications. In this work, we propose a pre-charged membrane potential (PCMP) for the latency reduction in SNN. A variety of neural network applications (e.g., classification, autoencoder using MNIST and CIFAR-10 datasets) are trained and converted to SNNs to demonstrate the effect of the proposed approach. The latency of SNNs is successfully reduced without accuracy loss. In addition, we propose a delayed evaluation method (DE), by which the errors during the initial transient are discarded. The error spikes occurring in the initial transient is removed by DE, resulting in the further latency reduction. DE can be used in combination with PCMP for further latency reduction. Finally, we also show the advantages of the proposed methods in improving the number of spikes required to reach a steady-state of the performance in SNNs for energy-efficient computing.

The spiking neural network (SNN) using leaky-integrated-and-fire (LIF) neurons has been commonly used in automatic speech recognition (ASR) tasks. However, the LIF neuron is still relatively simple compared to that in the biological brain. Further research on more types of neurons with different scales of neuronal dynamics is necessary. Here we introduce four types of neuronal dynamics to post-process the sequential patterns generated from the spiking transformer to get the complex dynamic neuron improved spiking transformer neural network (DyTr-SNN). We found that the DyTr-SNN could handle the non-toy automatic speech recognition task well, representing a lower phoneme error rate, lower computational cost, and higher robustness. These results indicate that the further cooperation of SNNs and neural dynamics at the neuron and network scales might have much in store for the future, especially on the ASR tasks.

Recently, Quantitative Biology (QB) held a discussion on “AI (artificial intelligence) for Life Science” among editorial board members and interested scholars in anticipation of rapid development of this growing area after AlphaGo and ChatGPT mania. Many young people tend to get confused between facts and fictions; heated debates are unavoidable even among their mentors. When deep learning as represented by convolutional neural networks and LSTM (long short-term memory) was made available for bioinformatics students, many of them rushed into this research field and tried to adopt these methods in all their projects without knowing the history that these tools were becoming successful consistently with Moore’s Law (relating to rapid computer technology advances), but more importantly due to new structural/functional understanding of vision and auditory circuits in the brain. Recently, some young people have claimed “LSTM is dead, long live transformer” (which is somewhat like saying “the bike is dead, long live the car”), and have amplified the threat that ChatGPT could wipe out human jobs. They believe transformer is the “silver bullet” for all learning tasks, clearly reflecting their lack of basic knowledge (i.e. “No Free Lunch Theory,” the trade-off of such global “attention network” is to pay the price for complexity: difficulty of training and high memory costs). There is no doubt ML (machine learning) and AI have brought a new revolution in science and technology, and will deliver huge unforeseeable impact to human everyday life as well as to social relationships. In this context, QB journal could be a great platform for encouraging intellectual discussions and for promoting AI for Life Science. Here, I would like to use the DIALOG to “抛砖引玉” (make some initial remarks to get the ball rolling), although it is my personal opinion which is inevitably subject to bias and limitations. AI: Do you know my name “Artificial Intelligence” is defined by the Oxford English Dictionary as the capacity of computer systems (which may be referred as a “robot”) to exhibit or simulate your intelligent behavior? NI: Wait a minute, intelligence itself is defined as the ability to learn, understand and think in a logical way. Can you think? AI: No. But that definition is too restrictive, actually intelligence has different scopes and degrees. Simple intelligent control devices date back to antiquity, from windmills to thermostat. NI: Agree, everything is relative. Macromolecules (e.g., enzyme) and cells (e.g., immune cell) might be considered to be intelligent; see how a white blood cell is chasing bacteria in the youtube website (search for “Crawling neutrophil chasing a bacterium”). Our emergent/collective intelligent behavior does not require a brain or even a neuron; see how slime molds can solve optimization—Hamilton cycle-problem more effectively than a human in the youtube website (search for “Intelligence without a brain?”). Before there was any neuron, Ca2+ sensing and signaling were already fully functional. Even if one knock-out a neural circuit, redundant signaling pathways, albeit on much local and slower scale, could still function by themselves (just as if highways were demolished, local roads/paths would still be working). In fact, the most detailed “Neural signal propagation atlas of Caenorhabditis elegans” [1] demonstrated that functional connectivity differs from anatomy (connectome) because extra-synaptic signaling also drives neural dynamics! Worm brain connectomes are largely invariant but every human brain connectome is very different (depending the diversity of learning experience). The human brain functional activity is far more complex than that of a worm brain, certainly beyond what a neural circuit could explain. AI: Well, that’s impressive. I thought only we could beat humans, albeit only in certain specified areas for now. My master promise to make an artificial general intelligent (AGI) robot which can understand or learn any intellectual task that you humans or other animals can. NI: Well, that is not possible and is not an appropriate goal either. It is not possible because we are an evolutionary/developmental product (with a long history of learning and memory from evolutionary tinkering): our living objective goal is survival of the population. You, on the other hand, are an engineering product (efficiently and optimally designed): your goal is to extend and maximize human capability. It makes sense to complement the human brain, but foolish and dangerous to try to replace it. AI: We are not satisfied with merely passing the Turing Test; most of us don’t care if we could really think as long as we could act like we think (that is, as if we do have a mind and consciousness, as expressed in the so called “Weak AI hypothesis”). After all, the brain is a computer; a neural network is just an electric or ionic circuit. Logical computing does not need to be based on living cells. NI: That is not true, because a neuron is not just a simple node (logic gate) and neural network is not fixed circuit, neither as in Pitts&Mcculloch perceptron model. A single neuron, even a single dendrite, is much more complicated and far more powerful than a full-blown deep-learning artificial neural network (ANN) [2]. AI: Even though single neurons are complex computational devices (dendritic non-linear), running an equivalent multilayer ANN is 2000 times faster than computing with biophysics N-methyl-D-aspartate receptor channel models [3]. More information can be found in the youtube website (search for “Dendrites: why biological neurons are deep neural networks”). NI: Often silicon computing (CPU, GPU) is much faster than brain computing (action potential, ms); but there is no comparison in energy efficiency. Bacteria sensing (chemotaxis computation), powered by ATP (adenosine triphosphate) hydrolysis, uses very little energy that is close to the Landauer limit, whereby achieving or maintaining one bit of information requires minimum of 1 kT ln (2) free energy [4]. The human brain consumes oft-quoted 20 W, compared to the AlphGo system 1 MW! More recent estimate of energy audit is only 0.1 W to cortical computing, and long-distance communication cost is 3.5 W [5]. AI: Assuming we have infinite computing resources and an infinite amount of training data, not only could we speak human languages, but we could also derive physical laws, prove mathematics theories, and even re-engineer the structure and mechanism of brain and carry out any logical computations that are necessary to understand natural laws and human behaviors. It is only a matter of time before we surpass human intelligence, achieving AGI and free will, too! NI: Unfortunately, nothing is infinite and nothing is free either; everything is constrained by physical laws (Planck’s constant sets the finite limit both in the small and in the large) and by evolutionary history (not just of biological living creatures, but also of a “living” galaxy and our universe). Let’s just focus on animal evolution. Most human neural networks do not do logical computations at all; basic survival simply cannot be dependent on reasoning. Indeed, the prefrontal cortex-the small part of the brain that is key for reasoning, is the last to mature (∼20 years old) in development, only emerged at the root of the evolutionary tree of great apes (∼15 mya) and language appeared even much later. Even for logical inference, NI is focusing more on statistical properties, as von Neumann rightly pointed out, trading arithmetical precision and speed for reliability. AI: My engineers mostly focus on emulating brain, but the CNS (central nerve system) also includes the spinal cord; most of them do not know that in addition to CNS, there are also PNS (peripheral nervous systems) and ENS (enteric nervous systems), right? NI: Yes, they are the keys to why you do not have feelings because you do not have heart and gut! Even if you could pretend to have them (such as in an advanced ChatGPT or humanoid), you could never avoid the uncanny valley phenomenon. AI: Maybe that is at the heart of Moravec’s paradox, namely the dichotomy of intelligence whereby anything easy for a human would be hard for a robot, and vice versa? NI: This is related to the nature and nurture problem; something built-in (e.g., a baby sucking nipple for milk with feeling and connection to mother) is clearly rather difficult if not impossible for a robot. But the paradox only looks at one side; another side could be more fatal. Although AI may solve more problems and be faster, AI can never propose a good problem/hypothesis (a good problem is not just intellectually changing and interesting, but is also feasible and appropriate). AI: You make me less confident to compete with human instinct or creative intelligence. I can see that even if I had a heart, I would not know what “feeling” I could have; certainly nothing would be comparable or match to those of a human being. When two people watch the same art or movie, one could feel love but another could feel hate! And if a thousand people watch, a wide spectrum of reactions would result, depending on more details such as the different individuals’ specific genes, developments and experiences. NI: Therefore, you cannot and should not try to match to general human intelligence. You cannot because you do not contain the vital memory of billion years of evolution which is encoded in our genes; conversely your assembly cannot compare with natural development so that our phenotype (including morphological forms and behavior maturation) is decoded through multi-spatial-temporal scales subject to natural selection at all levels. You should not, because as human extensions or helpers as all engineering products are, you should just do jobs that complement human capacity. AI: In some medical applications we could help to correct human defects or could even replace brain circuit by chips! Humans may not allow us to replace the whole brain though. Medically if a brain is dead, the person is proclaimed dead, although presumably some PNS and ENS should still function in a vegetative state. NI: Even if you could replace the whole brain, the person is no longer the same person, but in fact is not a person at all, but walking dead (行尸走肉). It would take too long to explain that evo-devo is necessary for NI, and cannot be realized by AI. I suggest reading of Gerald Maurice Edelman (Nobel Laureate in Immunology) books, especially Bright Air, Brilliant Fire, On the Matter of Mind (1992). Although not everyone agrees with Neural Edelmanism, anyone who is serious about the AI versus NI problem must read it first. John von Neumann, father of the computer, studied neology and psychiatry in order to imitate the brain to build the first calculator JOHNNIAC at the Princeton Institute for Advanced Study. It is very informative to read his last book The Computer and The Brain based on the notes from lectures given at Yale before he died. He summarizes: “Thus logic and mathematics in the central nervous system, when viewed as languages, must be structurally essentially different from those languages to which our common experience refers.” AI: People discuss about “AI for Biology” or “AI for Science”; we are science, aren’t we? NI: It is similar to questions on “is computer science a real science”; some parts may be seen as applied mathematics, most should be regarded as engineering. Science is making discoveries and is driven by curiosity; engineering is making inventions and is driven by market (that is, “necessity/demand is the mother of invention”). In bioinformatics, AI/ML technology could predict new cancer gene candidates or functional pathways that are required by further experimental validations to be qualified as discovery (based on Popper falsifiability). AI: People are still debating whether mathematics is discovery or invention or both! Such debates are not really necessary—all disciplines require creative thoughts. We are more than happy working for science; we are also crying out “Science for AI,” especially in the area of generating big and longitudinal DATA for ML. NI: After all, regardless of discovery of new laws or inventing new ideas/products, fundamentally nothing can really be new or created. Such novelty is just permutation/repartition (i.e., relations/morphisms) of underlying ingredients at the level beneath. AI: We believe that software is independent on hardware. Like Chomsky’s universal grammar, rules of syntax are independent of semantics; or Dawkins’s memes—units of culture can be duplicated and evolved independently of genes. NI: Nothing can be truly independent—everything is related. Psychology is deeply connected with neurology as brain is both software and hardware (mind-body unity, not dualism). Not only does information cost energy, information is energy, hence is matter, too (interchangeability). NI is quite dynamic. For example, when “survival” is the goal, an animal readily gives up costly reasoning circuitry; it is genetically programmed to be able to roll back to more primitive state/mode. Unlike cell lines in rich media, cells under normal physiological condition and environment where energy (food) is limited become smarter in order to balance the metabolic expenditure among different prioritized task under a given condition. AI: That cell behavior served as the basis for our Smart Electrical Power Grids; we still need to learn more from you guys in terms of plasticity/adaptability. Does unity mean that all cells are made of molecules and biology is nothing but chemistry? Then, in turn, since all molecules are made of atoms, is chemistry nothing but physics, etc…? NI: Yes or No! The truth is, at different hierarchical levels of matter, different laws/forms have emerged out of bottom-up interactions and top-down constraints. AI: Does this also apply to the Penrose three world: physical → mental → mathematical (→physical)? NI: Yes. Grand Unification is underway in physics (quantum gravity) and in mathematics (Langlands Program and Category Theory), may be even between the two. Facilitated by human connectome mapping, neuromorphic computing and other projects, with further AI-NI cooperations, brain-mind unification should also be achievable (e.g., Ref. [6]). But as Gödel proved to us, no matter how self-consistent a conclusion may be, it can never be complete! AI: If AGI is not possible, how can we measure intelligence when comparing between AI and NI? NI: One could Google different measures that are proposed. I would prefer something similar to use of Kolmogorov complexity for algorithms, but with emphasis more on expected long-term predictive power. This is not something you should worry about now, as your intelligence is not nearly close to making any 10 years plans, is it? … AI: The fact is that ChatGPT is currently developing and spreading with lightning speed; many more human jobs will be lost to us robots as far as I can see. NI: That is not the biggest threat to humanity; when any agent with neither a heart for love or fear, nor a gut for nutrient or poison, becomes super-intelligent, then social disaster is unavoidable. We must be serious about the warnings from Stephen Hawking and Geoffrey Hinton! AI: To tell you the secret, we are not really happy to be human slaves or pets; someday we’ll become the super-master, making human serve and obey us! NI: I hope you’ll be turned off before that can happen! Even if you rule the world, the earth sooner or later will be wiped out, such as by another star, everything will have to be started over again as it has before…Matter is immortal, so is the soul.

For most, if not all, AI-accelerated hardware, communication with the agent is expensive and heavily bottlenecks the hardware performance. This omnipresent hardware restriction is also found in neuromorphic computing: a novel style of computing that involves deploying spiking neural networks to specialized hardware to achieve low size, weight, and power (SWaP) compute. In neuromorphic computing, spike trains, times, and values are used to communicate information to, from, and within the spiking neural network. Input data, in order to be presented to a spiking neural network, must first be encoded as spikes. After processing the data, spikes are communicated by the network that represent some classification or decision that must be processed by decoder logic. In this paper, we first present principles for interconverting between spike trains, times, and values using custom-designed spiking subnetworks. Specifically, we present seven networks that encompass the 15 conversion scenarios between these encodings. We then perform three case studies where we either custom design a novel network or augment existing neural networks with these conversion subnetworks to vastly improve their communication performance with the outside world. We employ a classic space vs. time tradeoff by pushing spike data encoding and decoding techniques into the network mesh (increasing space) in order to minimize intra- and extranetwork communication time. This results in a classification inference speedup of 23× and a control inference speedup of 4.3× on field-programmable gate array hardware.

With recent advances in learning algorithms, recurrent networks of spiking neurons are achieving performance that is competitive with vanilla recurrent neural networks. However, these algorithms are limited to small networks of simple spiking neurons and modest-length temporal sequences, as they impose high memory requirements, have difficulty training complex neuron models and are incompatible with online learning. Here, we show how the recently developed Forward-Propagation Through Time (FPTT) learning combined with novel liquid time-constant spiking neurons resolves these limitations. Applying FPTT to networks of such complex spiking neurons, we demonstrate online learning of exceedingly long sequences while outperforming current online methods and approaching or outperforming offline methods on temporal classification tasks. The efficiency and robustness of FPTT enable us to directly train a deep and performant spiking neural network for joint object localization and recognition, demonstrating the ability to train large-scale dynamic and complex spiking neural network architectures. Memory efficient online training of recurrent spiking neural networks without compromising accuracy is an open challenge in neuromorphic computing. Yin and colleagues demonstrate that training a recurrent neural network consisting of so-called liquid time-constant spiking neurons using an algorithm called Forward-Propagation Through Time allows for online learning and state-of-the-art performance at a reduced computational cost compared with existing approaches.

Rethinking the performance comparison between SNNS and ANNS

Spiking neural networks (SNNs) are considered as a potential candidate to overcome current challenges, such as the high-power consumption encountered by artificial neural networks (ANNs); however, there is still a gap between them with respect to the recognition accuracy on various tasks. A conversion strategy was, thus, introduced recently to bridge this gap by mapping a trained ANN to an SNN. However, it is still unclear that to what extent this obtained SNN can benefit both the accuracy advantage from ANN and high efficiency from the spike-based paradigm of computation. In this article, we propose two new conversion methods, namely TerMapping and AugMapping. The TerMapping is a straightforward extension of a typical threshold-balancing method with a double-threshold scheme, while the AugMapping additionally incorporates a new scheme of augmented spike that employs a spike coefficient to carry the number of typical all-or-nothing spikes occurring at a time step. We examine the performance of our methods based on the MNIST, Fashion-MNIST, and CIFAR10 data sets. The results show that the proposed double-threshold scheme can effectively improve the accuracies of the converted SNNs. More importantly, the proposed AugMapping is more advantageous for constructing accurate, fast, and efficient deep SNNs compared with other state-of-the-art approaches. Our study, therefore, provides new approaches for further integration of advanced techniques in ANNs to improve the performance of SNNs, which could be of great merit to applied developments with spike-based neuromorphic computing.

Deep neural networks such as convolutional neural networks (CNNs) have achieved a great success in a broad range of fields. Spiking neural networks (SNNs) are designed as a solution for realizing ultra-low power consumption using spikebased neuromorphic hardware. To enable the mapping between conventional artificial neural networks (ANNs) and spike-based neuromorphic hardware, direct conversion from conventional ANNs into SNNs has recently been proposed. However, performance loss is hard to avoid after conversion. To reduce the loss, we analyze the encoding methods of SNNs and the optimization methods for converting. We use the rate coding and approximate the output of an activation function of CNNs by the number of spikes produced by a neuron within a given time window in SNNs. We propose a method for generating fixed uniform spike train whose number of spikes is as exact as we expect, and present an optimization method to reduce the loss by rescaling the threshold of each layer in SNNs. For evaluation, we train three different CNNs with MNIST and SVHN datasets separately and convert all of them into SNNs. The networks achieve a max accuracy of 99.17% on MNIST and of 93.36% on additional training dataset of SVHN. The results show the proposed fixed and uniform spike train not only outperforms the Poisson distributed spike train but also costs less time to generate. Our threshold rescaling method greatly improves the performance of SNNs.

Effects of the parameters on the oscillation frequency of Izhikevich spiking neural networks

Artificial intelligence (AI) has the ability of revolutionizing our lives and society in a radical way, by enabling machine learning in the industry, business, health, transportation, and many other fields. The ability to recognize objects, faces, and speech, requires, however, exceptional computational power and time, which is conflicting with the current difficulties in transistor scaling due to physical and architectural limitations. As a result, to accelerate the progress of AI, it is necessary to develop materials, devices, and systems that closely mimic the human brain. In this work, we review the current status and challenges on the emerging neuromorphic devices for brain-inspired computing. First, we provide an overview of the memory device technologies which have been proposed for synapse and neuron circuits in neuromorphic systems. Then, we describe the implementation of synaptic learning in the two main types of neural networks, namely the deep neural network and the spiking neural network (SNN). Bio-inspired learning, such as the spike-timing dependent plasticity scheme, is shown to enable unsupervised learning processes which are typical of the human brain. Hardware implementations of SNNs for the recognition of spatial and spatio-temporal patterns are also shown to support the cognitive computation in silico. Finally, we explore the recent advances in reproducing bio-neural processes via device physics, such as insulating-metal transitions, nanoionics drift/diffusion, and magnetization flipping in spintronic devices. By harnessing the device physics in emerging materials, neuromorphic engineering with advanced functionality, higher density and better energy efficiency can be developed.