Taking inspiration from the human oculomotor network, Polykretis et al. designed a SNN crafted to perform real-time tracking of visual targets [23]. The designed oculomotor system was implemented to command an in-house robotic head equipped with two eyes and a neck. The brain-constrained controller was deployed on Intel’s neuromorphic chip, demonstrating its capability to track slow- and fast-moving visual targets in real-time without the need for training. They utilized Intel’s neuromorphic chip to implement the SNN controller with brain-constrained characteristics. The demonstration highlighted its real-time tracking capability for slow- and fast-moving visual targets, achieved without the need for training. The observed behavior closely resembled the smooth pursuit and saccadic eye movements seen in the human oculomotor system. Furthermore, the SNN’s performance was comparable to that of a CPU-operated PID controller. To elaborate, the robotic head boasts 6 degrees-of-freedom (DOF), with 2 DOF assigned to the neck and 2 DOF for each of the two eyes.

The robot oculomotor system is divided into three main parts [23]. Firstly, there is the biomimetic robotic head, which plays a crucial role in tracking moving objects in the environment. Next, there is the robot operating system, responsible for controlling the robot’s movements. The most essential and final component is the oculomotor SNN, comprising an encoder and Loihi processor. The oculomotor SNN was executed on Intel’s Loihi neuromorphic processor, employing 74 compartments and 78 connections on a single Neurocore. This configuration facilitated real-time control of the biomimetic robotic head through the interaction framework. In collaboration, the controller and Loihi played roles in visual input processing, omnidirectional 1 degree-of-freedom (DOF) control, and head control in this study. As a result, it can be observed that except for cases with very high speed or short distance of the object tracking, the robot head properly tracks the movement of objects.

Yan et al. introduce a bionic olfactory signal processing (BOSP) network utilizing a SNN to tackle related challenges [24]. Modeled after the mammalian olfactory bulb structure, the system comprises two components: a virtual olfactory receptor (VOR) layer and a bionic olfactory bulb (BOB) layer. The BOSP network emulates the biological olfactory system, incorporating VORs to capture sensor array signals and BOBs to simulate olfactory bulb processing, ultimately generating features for odor recognition by classifiers. The VOR layer in the BOSP network employs the principles of SNN for spike coding. It transforms the physical and chemical information of various odors into distinguishable electrical spike sequences using SNN encoding methods. The BOSP network operates at the feature level, automatically extracting valuable features from sensor array outputs, eliminating the need for traditional and intricate feature extraction steps. The e-nose system encompasses three distinct procedures with varied structures and functions: (1) olfactory perception, simulating olfactory receptors receiving odor signals through a sensor array module; (2) olfactory conduct, simulating olfactory bulb processing of olfactory signals through a data processing module; and (3) olfactory recognition, simulating the identification of olfactory signals by the piriform cortex through a pattern recognition module.

To implement SNN, the LIF model was adopted, and the Spike-Timing-Dependent Plasticity (STDP) learning rule was employed [24]. The pivotal component in this setup is the bionic olfactory bulb, which introduces competition among neurons in the network and determines classification outcomes based on the activation levels of the neurons. The system detects four categories of wounds in Sprague Dawley rats using a gas sensor array: uninfected wounds, and wounds infected with three different bacteria—Staphylococcus aureus, Escherichia coli, and Pseudomonas aeruginosa.

The BOSP model exhibits superior performance across all evaluation indices compared to other studies, signifying its stable ability to distinguish each type of gas [24]. Notably, the BOSP model demonstrates significantly higher sensitivity and precision, suggesting its increased likelihood to accurately identify the presence of an odor. It was an innovative challenge to distinguish odors. Seeing the effectiveness of SNN even in such novel areas indicates the high applicability of SNN in the broader field of biomedicine.

Covid-19 has emerged as a significant global health concern, necessitating early detection for effective treatment and prevention of further transmission. Kamal et al. introduce a SNN employing supervised synaptic learning for the identification of abnormalities in Chest X-rays (CXRs) [25]. Specifically, the proposed SNN is designed to differentiate between Covid-19 positive cases and healthy individuals. The utilization of SNN is pivotal in examining abnormalities, such as Covid-19, in CXRs due to its suitability for handling noisy data.

The proposed method employs the Izhikevich neuron model, chosen for its superior performance in both biological plausibility and computational efficiency, particularly in the context of large-scale SNN [25]. Synaptic weights from the hidden layer to the output layer are trained using spike-based supervised learning rules, specifically the tempotron learning procedure. Subsequently, Support Vector Machine (SVM), a supervised model based on statistical approaches for classification and regression, is utilized. The image database comprises three distinct groups, namely Covid-19, normal, and Viral Pneumonia, with 1200 X-ray images for Covid-19, 1341 images for normal, and 1345 images for Viral Pneumonia. Each image is in grayscale, with dimensions of 256 × 256 pixels, predominantly focusing on two groups of images: Covid-19 and normal.

The study conducted a comparison of their proposed method with two alternative prediction methods, KNN and the decision tree [25]. The results indicate that their approach achieves a higher accuracy rate of 78%, surpassing the accuracy of the other two methods, which stand at 65% and 67%, respectively. Despite this, the research does not exhibit a significantly superior accuracy when compared to other existing models. As a result, the practical commercialization potential of this technology remains uncertain. Additionally, the study suggests a need to address the challenge of unbalanced data in their analysis. The future research goals will involve enhancing accuracy and assessing the potential for commercialization, aiming to determine the feasibility of practical implementation.

Energy efficiency is a crucial factor when deploying deep neural networks (DNNs) at the edge, where power constraints may limit budgets to milliwatts or microwatts for inference tasks. Research in low-power neural network architectures is essential to gain insights for energy-constrained applications. Neuromorphic processors, designed to mimic the energy-efficient characteristics of the human brain, present a promising solution. Intel’s neuromorphic research chip, Loihi, designed for SNNs, offers new possibilities for assessment. Buettner et al. aim to assess the effectiveness of a leading SNN design approach, the artificial-to-spiking neural network (ANN-to-SNN) conversion, in combination with the innovative neuromorphic processor Loihi, to deliver an accurate and energy-efficient solution for heartbeat classification [28]. Key steps of ANN- to-SNN include translating neuron models, mapping activation functions to spike generation, quantizing weights, encoding input into spikes, considering temporal dynamics, translating learning rules, adapting network architecture, and conducting thorough testing and fine-tuning.

For the SNN-Toolbox study, a 1D-CNN architecture is selected [28]. The model is crafted to achieve high accuracy, possess ample size for meaningful benchmarks, and maintain a moderate number of layers to avoid substantial effects on the SNN-Toolbox conversion process. The training involves 50 epochs using TensorFlow 2.2.0, and following training, the model is transformed into an SNN with the same architecture using the SNN-Toolbox.

As a result, the SNN used in this study achieves an accuracy of 97.8% and a macro-averaged F1 score of 87.9% across five classes, slightly lower than the ANN’s performance of 98.4% accuracy and 90.8% F1 score [28]. The macro-averaged F1 score is a metric used for evaluating the performance of a classification model. It is calculated by computing the F1 score for each class individually and then averaging these scores across all classes.

Concerning performance metrics, it’s noted that Loihi demonstrates the least dynamic power consumption but concurrently registers the highest latency [28]. Notably, there’s a substantial need for latency improvements in Loihi. Additionally, given the lower accuracy of SNN when compared to ANN, it becomes apparent that there is room for improvements in accuracy as well. These areas of improvement can be regarded as potential directions for future work.

The extensive adoption of reinforcement learning (RL) has empowered robots to acquire complex skills through a process of trial and error, leveraging rewards and penalties. RL, often in tandem with CNNs, has showcased remarkable efficacy, especially in scenarios involving robots with numerous degrees of freedom (DoF) navigating continuous action spaces [30]. Despite the advanced capabilities of DNNs, they are hampered by significant power consumption. Consequently, there is a growing consensus that a synergistic integration of SNNs and DNNs holds promise in overcoming challenges faced by future robotic systems.

Oikonomou et al. provide distinct phases: Initially, a pioneering hybrid- deep deterministic policy gradient (DDPG) agent, incorporating both a spiking actor and a deep critic network, was introduced for the first time, operating on a 6-degree-of-freedom robotic arm within a 3-D action space [30]. Subsequently, an object detection algorithm was employed, enabling the spiking RL agent to autonomously explore in response to identified targets. Ultimately, an enhanced performance in the target reach task was observed when compared to the conventional DDPG approach. The key emphasis lies on the spiking actor network, which utilizes the LIF model. The training algorithm employed for this network is the STBP algorithm. To evaluate the performance of the hybrid-DDPG, a dedicated evaluation environment mirroring the training setup was created. For the evaluation phase, five random initial positions were generated, and the success rate and execution time were measured for each trial.

Certainly, since the DDPG algorithm is crafted to address high-dimensional challenges, potentially it gained advantages from a sub-goal learning structure [30]. Nonetheless, it is not predetermined that a sub-goals approach is invariably more effective than an end-to-end strategy, as it could elevate the complexity of the environment and demand additional training. Moreover, it is evident that employing the newly proposed Hybrid DDPG, as opposed to the classic DDPG, results in a higher success rate. Additionally, observing the reduced execution time suggests a potential decrease in power consumption. Through this, it can be inferred that the novel Hybrid DDPG approach utilizing the SNN has enhanced performance compared to the conventional method.