Neuromorphic Computing for AI: Bridging Artificial and Biological Neural Networks Through Spiking Models

I. Introduction

Neuromorphic computing is an area of research inspired by the architecture and functioning of biological neural networks. It tries to bridge the gap between artificial intelligence (AI) and the very complex processes of the human brain. Traditional AI models, such as deep learning, rely on artificial neural networks but fail to mimic the efficiency, flexibility and energy effectiveness of biological systems. Neuromorphic computing, therefore, addresses these limitations through the design of hardware and software systems that replicate the brain’s structure and its adaptive learning processes [1]. The human brain processes a vast amount of information in parallel and is highly energy efficient, yet surprisingly fast. Nevertheless, most machine learning models, even with the current major advances in AI, still rely on serial computation and require substantial computational resources and large amounts of energy [2]. In this context, neuromorphic computing tries to address such issues by incorporating the most critical neurobiological principles, namely, spiking neural networks (SNNs) and synaptic plasticity [3]. This intersection of AI and neuroscience could be a revolution in different fields such as robotics or healthcare, enhancing algorithms in machine learning with the best efficiency, adaptability and cognitive-like behavior. It would then become evident that the incorporation of neuromorphic computing can bring out a much stronger improvement in AI’s capability to work with complex dynamic environments and real-time learning. Processing information like biological systems could be opened up to even more natural, responsive and intelligent systems that can learn from experience and better relate to the world around them. The idea behind this study is to explore whether this innovation might close the gap between artificial and biological neural networks by exploring in-depth its implications and challenges for future developments in AI systems. It helps us find new understanding of and use for neuromorphic systems within practical applications in real-life conditions. AI has witnessed rapid growth in both research and application over the past 3 decades. This is evidenced by the sharp rise in AI-related publications, particularly after 2010, as shown in Figure 1.

Figure 1:

II. Literature Review

One of the important contributions has been made by Li et al. [4], proposing a spatiotemporal pruning method in training ultralow-latency SNNs. They have targeted remote sensing scene classification tasks using benchmark remote sensing datasets, where processing speed and low-latency inference are critical performance requirements. The proposed model optimizes the resources by not compromising the accuracy to achieve real-time classification tasks under complex environments. Typical evaluation metrics in such remote sensing neuromorphic classification studies include classification accuracy, inference latency, precision, recall and computational efficiency under real-time processing conditions. Similarly, Li et al. [5] researched the biomimetic neuromorphic sensory systems based on the integration of electrolyte-gated transistors. These mimic the biological sensory processes and, therefore, enhance energy efficiency and stability during operation, making them a good choice for wearable and portable devices. It highlights the possibility of using neuromorphic computing in cases where power and real-time processing constraints are of concern. Liu et al. [6] presented an elaborate road map of two-dimensional (2D) materials and integration of these into neuromorphic devices. Such a study makes clear the prospect in these applications by improving their functionality with neuromorphic chips containing graphene and other transition-metal dichalcogenide materials. This has provided some hope to transform the speed-processing power consumption level by making these neuromorphic computations have an easy key or enabling the processing to make the hardware necessary.

On the application side, Long et al. [7] developed a neuromorphic bionic eye that uses filter-free color vision based on hemispherical perovskite nanowire arrays. This work is an important step toward the integration of neuromorphic systems with advanced sensor technologies to build bioinspired vision systems. The integration of such neuromorphic systems in bionic devices can significantly improve sensory applications, such as vision prosthetics and assistive technologies.

Lakshmi Varshika et al. [8] are focused on the nonvolatile memories within the SNNs. Their research looks at the present and future trends of memristive devices applied for the emulation of synaptic behavior. They have, therefore, explored the nonvolatile memory role in the neuromorphic architectures by emphasizing its possible application for further improvement of the performance efficiency of SNNs, mainly in those systems where long retention of memory is needed. Recent work by Maldonado et al. [9] has explored memristive devices for neuromorphic computing. These devices, based on titanium nitride (TiN), titanium (Ti) and hafnium oxide (HfO₂), are known to replicate synaptic plasticity and stochastic resonance phenomena. The authors demonstrate how memristive systems can be used to model brain-like learning behaviors, thus leading to more robust and adaptive neuromorphic systems. Further, Nilsson et al. [10] discussed the integration of neuromorphic computing with real-time data processing systems by exploring event-driven distributed systems for brain-inspired computing. The work has shown how the neuromorphic systems process data in an asynchronous event-driven manner, similar to the human brain, to improve computational speed and energy efficiency. The application of neuromorphic computing in health systems has also been taken very seriously. Pedersen et al. [11] proposed a unified instruction set for neuromorphic computing to facilitate interoperability between various brain-inspired computing systems. This would imply tremendous possibilities for integrating neuromorphic technologies into medical devices and healthcare applications to realize more accurate and efficient diagnostics.

Unlike conventional von Neumann systems, where memory and processing are separated, neuromorphic architectures integrate these functions, significantly reducing data movement and energy consumption as illustrated in Figure 2.

Figure 2:

Recent developments in neuromorphic computing have contributed to richer theoretical and practical insights into brain-inspired architectures. Although earlier studies have concentrated on recreating neuronal firing via mimicking, recent efforts have focused on the integration of the system, device invention, scalability and cross-domain application. The development of large-scale neuromorphic chips such as the Loihi architecture by Intel can be considered among the most impactful advances in neuromorphic hardware. As an example, Davies et al. [12] proposed Loihi, a many-core neuromorphic processor that combines on-chip learning and programmable mechanisms of synaptic plasticity. Loihi showed a significant level of energy efficiency relative to graphics processing unit (GPU)-based systems, especially with constraint satisfaction and optimization problems. Future research also confirmed its usability in robotics and adaptive control systems. The Loihi architecture primarily supports local on-chip learning mechanisms and programmable synaptic plasticity, while surrogate-gradient-based SNN frameworks commonly employ supervised optimization strategies. Several recent neuromorphic systems additionally combine local spike-driven adaptation with global gradient-based optimization to form hybrid learning architectures [12].

In the same way, the TrueNorth chip by international business machines (IBM) was a breakthrough in neuromorphic event-based architecture. Merolla et al. [13] developed a million-neuron architecture, which was able to run ultralow-power pattern recognition. In contrast to the conventional deep learning accelerators, the TrueNorth used asynchronous spike-based communication and delivered real-time inference at minimal energy cost. In addition to digital neuromorphic processors, analog and mixed-signal neuromorphic circuits have also been of interest. A recent paper by Indiveri and Liu [14] has examined analog VLSI implementations of SNNs and their potential benefits in low-latency sensory applications. The analog neuromorphic circuits minimize quantization errors and enable continuous-time dynamics, which are similar to biological neurons.

In materials, studies on resistive random-access memory devices have progressed greatly. Prezioso et al. [15] achieved neural networks based on memristors that were able to complete pattern recognition tasks with the help of crossbar arrays. The crossbar architectures allow in-memory matrix–vector multiplication, which lowers the computational latency and power consumption significantly. This type of innovation is particularly relevant to the present study because it supports the implementation of memristive synaptic plasticity in neuromorphic computing systems. Moreover, van de Burgt et al. [16] created organic electrochemical transistors, which can reproduce the effect of a synapse. Such devices enable both short-term and long-term plasticity, which is appropriate in flexible and wearable neuromorphic systems. Their ability to work in a biocompatible manner opens the possibilities of medical implantable neuromorphic devices. Table 1 presents a comparative summary adapted from the study’s report.

Surrogate gradient learning has become the new wave of technology on the algorithmic front in terms of training deep SNNs. Neftci et al. [17] suggested surrogate gradient techniques that provide an approximation of the non-differentiable spike function and thus allow training SNNs with backpropagation. This innovation tries to deal with one of the largest restrictions of neuromorphic learning, which is training instability. Bellec et al. [18] also implemented long short-term memory (LSTM)-like dynamics in spiking neurons, which enhanced credit assignment in temporal recurrent spiking networks. Their results indicate that a set of biologically inspired recurrent dynamics is capable of competing with traditional recurrent neural networks (RNNs) in sequence learning processes.

Neuromorphic vision sensing is another direction that is important. Lichtsteiner et al. [19] came up with the dynamic vision sensor, an event-based camera, which records variations of luminance instead of complete frames. The use of event-based sensors and SNNs makes the redundancy of data minimal and allows motion detection at high speed. More recent efforts by Gallego et al. [20] surveyed event-based vision algorithms and showed that they can be better used in low-latency robotics. Neuromorphic computing has also been used more frequently for brain-to-machine interfaces (BMIs). It has been shown that Pandarinath et al. [21] achieved high-performance neural decoding with adaptive recurrent architectures. Implementation of neuromorphic processors into BMIs has the potential of offering energy-efficient and implantable neural decoding methods.

Recent studies in the field have been directed at the explanation of spike-based decisions. Spike-based saliency mapping methods were suggested by Kim and Panda [22] to achieve better interpretability of SNN predictions. Elucidable neuromorphic AI is essential in healthcare and defense use where transparency is a requirement. Moreover, neuromorphic reinforcement learning has also been on the upswing. Strengthening reward-modulated spike timing-dependent plasticity (STDP) reinforced learning in spiking networks was proposed by Fremaux and Gerstner [23]. The strategy brings together biological dopamine-based learning systems and computational reward systems. Another emerging field is the integration of edge computing. Blouw et al. [24] tested the idea of neuromorphic systems with real-time keyword spotting and showed orders-of-magnitude energy efficiency improvements over deep neural networks. Their activity justifies the use of neuromorphic in the Internet of Things (IoT) and embedded AI.

Security of neuromorphic hardware is also becoming a new research direction. Shukla et al. [25] researched the vulnerability of memristor-based neuromorphic systems to hardware attacks, but they did not find any side-channel attack vectors. With the progress of neuromorphic chips toward commercial implementation, security on the hardware level has to be given high priority. Moreover, there are experimental studies of hybrid neurosymbolic methods to combine symbolic reasoning with spike-based learning. Garcez et al. [26] talked about neurosymbolic AI models that utilize neural architecture with logical reasoning. Adding such mechanisms to neuromorphic systems can establish reasoning better than pattern recognition. Finally, the issue of sustainability has become more pertinent. According to Strubell et al. [27], the environmental cost of large-scale deep learning models was emphasized. Neuromorphic computing provides an alternative that is highly promising due to its much lower energy use in training and inference, as well as meets the goal of sustainable AI.

Taken together, these papers support the increasing maturity of neuromorphic computing hardware, algorithms, applications and interdisciplinary research. Although traditional deep learning still sets the pace in large-scale benchmarks, neuromorphic systems have distinct benefits in the areas of energy efficiency, real-time flexibility and biological plausibility. The paradigm is a promising approach in the future generation of intelligent systems, which integrates novel materials, event cameras, surrogate gradient training and neuromorphic scale processors.

Intelligent distributed AI systems have also made significant strides in recent years, which have helped them achieve adaptive and efficient computing architectures. To achieve effective human activity recognition in dynamic sensing environments, Raja Sekaran et al. [28] introduced the hybrid spatial-temporal convolutional network–nu-support vector classifier (HSTCN-NuSVC) deep ensemble framework for human activity recognition using smartphone sensors.

In recent years, the research has also delved into adaptive intelligent optimization frameworks for AI systems. In dynamic operational settings, intelligent decision-making is critical for optimal vehicle maintenance, as suggested by Meng [29], who introduced an adaptive scheduling framework based on deep reinforcement learning for vehicle maintenance optimization. In addition, Xiao and Dong [30] studied optimization methods for applying artificial intelligence generated content (AIGC) technologies to IoT devices, focusing on computational efficiency and intelligent edge deployment, with a deep learning approach. These are complementary studies that are relevant to scalable neuromorphic and edge-AI architectures.

III. Methodology

Neuromorphic computing and AI research is a comprehensive search of the data, algorithms and methods involved in analyzing the interactions between biological neural networks and AI. The data adopted for the training and evaluation of the models are presented in the next paragraph, and subsequently, the four main algorithms adopted in our study are described. To capture the functionality and efficiency of these computational techniques, we also give a few tables of key results and pseudocode for each algorithm, which try to capture the functionality and efficiency of the computational techniques [31].

The methodological framework includes comparative evaluation of SNN, STDP, Hebbian learning and RNN models using benchmark datasets such as modified national institute of standards and technology database (MNIST), Canadian institute for advanced research 10-class dataset (CIFAR-10) and synthetic spike data. Performance comparison was conducted using accuracy, precision, recall, training time, memory utilization and latency-related evaluation criteria.

b. Algorithms

This research on neuromorphic computing focused on four algorithms. The algorithms were carefully chosen due to their importance to neural network modeling and a good potential for bridging the gap between artificial and biological systems. The following is a discussion of these algorithms.

1). SNN

SNNs are a type of artificial neural network that is designed to mimic the way that neurons communicate with each other in the brain by sending signals in the form of spikes. This operation differs from the non-spiking variant in that it considers the temporal dynamics of the operation but only emits a spike when the membrane potential is above a threshold value [34]. In this respect, SNNs act in a more realistic fashion than real neurons. Biological neuronal behavior is computationally modeled using membrane potential integration, threshold-triggered spike generation and synaptic weight adaptation mechanisms. The leaky integrate-and-fire (LIF) neuron model emulates temporal charge accumulation in biological neurons, while STDP-based learning reproduces activity-dependent synaptic modification observed in biological neural systems.

• Functionality: SNNs are created to be information processing devices in time. Unlike conventional neural networks, which have the property of processing information through layers, the SNN conveys spikes and computes with them by using their timing and frequency.

• STDP: Here, synaptic weights are dependent on the relative timing between the spikes of the two neurons, pre- and postsynaptic neurons.

• LIF Model: A model of neurons that integrates spikes and “leaks” over time; a neuron fires when its membrane potential crosses the threshold. The dynamics of the membrane potential in a spiking neuron can be expressed as:

  (1)

  $${\tau }_m\frac{dV\left(t\right)}{dt}=-\left(V\left(t\right)-V_{rest}\right)+RI\left(t\right)$$
  where V(t) is the membrane potential at time t, rest V_rest is the resting potential, τ_m is the membrane time constant, R is resistance and I(t) is the input current. A spike is generated when V(t) exceeds a threshold V_th, after which the membrane potential is reset.

SNNs represent a fundamental component of neuromorphic systems by simulating brain-like spike-based signaling between neurons, as shown in Figure 3.

Figure 3:

Algorithm 1: Spiking Neural Network Simulation

• 1: Initialize network with N neurons and synaptic connections

• 2: for each timestep do

• 3:  for each neuron do

• 4: Calculate membrane potential based on incoming spikes

• 5: if membrane potential exceeds threshold then

• 6:   Fire spike and reset membrane potential

• 7:  Adjust synaptic weights

• 8: end if

• 9:  end for

• 10:  Update synaptic weights using STDP

• 11: end for

2). STDP

STDP is the learning rule based on the temporal relation between pre- and postsynaptic spikes. One of the primary biological mechanisms of learning, it has been applied successfully to train SNNs in neuromorphic computing.

• Functionality: The rule varies the strength of the synaptic connections with respect to whether the presynaptic neuron is spiking before or after the postsynaptic neuron. If the pre-neuron is spiking right before the post-neuron, the connection will be strengthened, and the connection will be weakened if the post-neuron spikes before the pre-neuron [35].

• Presynaptic Spiking: The more in advance the presynaptic neuron spikes compared to the postsynaptic neuron, the greater the increase in the synaptic weight.

• Postsynaptic Spiking: If the postsynaptic neuron spikes before the presynaptic neuron, synaptic weights are decreased.

Algorithm 2: STDP Weight Update Rule

• 1: for each synapse do

• 2:  if pre-synaptic spike occurs before post-synaptic spike then

• 3:   Increase synaptic weight

• 4:   else

• 5:   Decrease synaptic weight

• 6:   end if

• 7: end for

The change in synaptic weight based on the relative timing between pre- and postsynaptic spikes is defined as:

(2)

$$\begin{array}{l}\Delta w=A_{+}exp\left(-\Delta t/{\tau }_{+}\right),\hspace{0.17em}if\hspace{0.17em}\Delta t>0,\hfill \\ \Delta w=-A^{-}exp\left(\Delta t/{\tau }^{-}\right),\hspace{0.17em}if\hspace{0.17em}\Delta t<0\hfill \end{array}$$

where Δt = t_post − t_pre

3). Hebbian Learning Rule

Hebbian learning is another biological rule applied in artificial neural networks as well as in neuromorphic systems. It can often be summarized as “cells that fire together, wire together,” meaning that the simultaneous activation of two neurons strengthens their connection. The weight update mechanism in Hebbian learning can be expressed as:

(3)

$$\Delta w_i{}_j=\eta \times x_i\times y_i$$

The Hebbian learning rule updates the synaptic weight based on the correlation between neuron activations. Here, Δw_Ij represents the change in synaptic weight between neuron i and neuron j,η is the learning rate, x_i is the presynaptic neuron activity and y_j is the postsynaptic neuron activity. The equation implies that the connection strength increases when both neurons are activated simultaneously.

• Functionality: The rule is the synaptic updating based on the correlation of the presynaptic and postsynaptic neurons. This increases when two neurons fire in more frequent cycles, so synapses become stronger for easy learning [36].

• Correlated Activity: The rule fortifies synapses between correlated neurons as far as firing patterns are concerned.

• Synaptic Weight Update: The synaptic weights are incremented when both pre- and post-neurons are fired together.

Algorithm 3: Hebbian Learning Rule

• 1: for each neuron pair do

• 2: if both neurons fire together then

• 3:  Increase synaptic weight

• 4: else

• 5:  Keep synaptic weight unchanged

• 6: end if

• 7: end for

4). RNN

RNNs are built with the idea of handling sequential data by keeping a state, which is updated at each time step. RNNs can utilize the information of previous decisions for the current decision, so it would be great for tasks such as speech recognition or time-series prediction [37].

• Functionality: The recurrent connections in the RNN ensure that memory traces for preceding inputs are retained; hence, the sequential data can be processed by RNNs with more precision than feedforward networks.

• Hidden State: Each time step calculates an updated internal state based on both the input and its previous state. The hidden state update in an RNN is defined as:

  $$h_t=tanh\left(W_h\times h_{\left(t-1\right)}+W_x\times x_t+b\right)$$

  (4)

  $$y_t=W_y\times h_t$$

The RNN updates its hidden state h_t at each time step using the current input x_t and the previous hidden state h_(t-1). The tanh activation function introduces nonlinearity, enabling the model to capture complex temporal dependencies. W_h and W_x. are weight matrices, and b is the bias term. The output y_t is computed from the hidden state using weight matrix W_x.

Vanishing Gradient Problem: RNNs often suffer from the gradient vanishing problem during back-propagation, but variants such as LSTMs have helped avoid this problem.

Algorithm 4: Recurrent Neural Network (RNN) Training

• 1: Initialize weights and biases

• 2: for each timestep do

• 3:  Calculate hidden state using the previous hidden state and input

• 4:  Compute output prediction from the current hidden state

• 5: Update weights using backpropagation

• 6: end for

IV. Experiments

d. Experiments on synthetic spike data

Training Setup for Synthetic Spike Data

• SNN: SNN was also trained using 500 neurons and with more complex interactions of the synapses [39]. The time intervals were indeed much smaller at 50 ms for better capture of the spikes’ temporal dynamics.

• STDP: It is the rule in spiking networks that aims for the association of stimuli.

• To see if it can learn patterns in the spiking data, a simple network was employed, called Hebbian learning.

• RNN: An RNN with LSTM cells was used to extract the long-term dependency information from the spike train data for training [40].

2). Results on Synthetic Spike Data

In Table 4, the accuracy, training time, memory usage, precision and recall are used as performance metrics, and the performance of the SNN, STDP, Hebbian learning and RNN models is compared on synthetic spike data. The accuracy of the SNN was the highest (92.0%), outperforming the RNN and other biologically inspired learning models. The result suggests that spike-based neural architectures work well for temporally structured spike train data and have lower memory consumption than RNN-based architectures. Due to its restricted learning complexity in more dynamic spike-driven situations, STDP and Hebbian learning demonstrated lower performance scores compared to other learning methods [41].

Table 4:

Comparative performance of learning algorithms on synthetic spike data

Algorithm	Acc. (%)	Train time (s)	Memory (MB)	Precision (%)	Recall (%)
SNN	92.0	180	150	91.5	92.3
STDP	85.4	210	160	74.8	75.3
Hebbian learning	80.5	190	130	79.7	80.9
RNN	88.2	240	250	87.5	88.9

[i] RNN, recurrent neural network; SNN, spiking neural network; STDP, spike timing dependent plasticity.

3). Analysis

• In this task, SNNs performed best and worked well on the synthetic spike dataset that was especially crafted to test the temporal dynamics of spike-based systems [41].

The RNNs were able to work well but slightly less accurately than the SNNs, likely due to the nature of the spike data being sequential, better suited to the spiking neurons than the LSTM Units.

Both STDP and Hebbian learning were impaired by data complexity and immature learning rules and would fare somewhat better.

A computing framework for the neuromorphic materials is shown in Figure 4, where the intrinsic physical properties of the materials, including phase transitions, ionic conductivity or memristive behaviors, are used to emulate neural processing. They allow for energy-efficient, parallel, adaptive computation in the hardware substrate itself.

Figure 4:

V. Discussion and Future Perspectives

VI. Conclusion

Neuromorphic computing is a paradigm shift in the field of finding a convergence between artificial and biological intelligence. Comparative analysis of the SNNs, STDP, Hebbian learning and RNN models shows that although traditional RNNs have reached a high classification accuracy, neuromorphic models have significant benefits in terms of energy efficiency, processing time and hardware alleviation. Scalability, as well as adaptability, is further promoted by the integration of memristive devices and new materials. Despite the obstacles associated with integrating goods on a large scale and the complexity of training, current studies indicate that neuromorphic computing will be central to the next generation of intelligent systems. The gap between artificial and biological neural paradigms is not just a technological innovation but also a stepping stone toward the realization of sustainable, adaptive and real-time intelligent computing systems.