Search

Artificial intelligence technologies based on deep convolutional neural networks and large language models have made significant breakthroughs in many tasks, such as image recognition, target detection, semantic segmentation, and natural language processing, but also face a conflict between the high computational capacity of the algorithms and limited deployment resources. Quantization, which converts floating-point neural networks into low-bit-width integer networks, is an important and essential technique for efficient deployment and cost reduction in edge computing. This paper analyzes various existing quantization methods, showcases the deployment accuracy of advanced techniques, and discusses the future challenges and trends in this domain.

We examine the impact of air pollution on underpricing of initial public offerings (IPOs) and its underlying mechanisms. Our findings suggest that IPO underpricing increases when air pollution is more severe in cities where bidding investors are located. Mechanism tests show that air pollution influences investors' mood and cognition, leading them to make pessimistic bids and rely more on anchors. Additional tests document that the impact of air pollution on IPO underpricing is more salient when information asymmetry is more severe and when the IPO firm also faces severe air pollution. Our research contributes to the literature on the causal influence of air pollution on pricing efficiency in primary markets by linking air pollution to potential bias embedded in investors’ bid prices, which carries important implications for environmental regulations in emerging markets, given the economic cost of air pollution imposed on IPO firms.

IntroductionThe complement system is a key component of the innate immune system, and its aberrant activation underlies the pathophysiology of various diseases. Zilucoplan is a macrocyclic peptide that binds and inhibits the cleavage/activation of human complement component 5 (C5). We present in vitro and ex vivo data on the mechanism of action of zilucoplan for the inhibition of C5 activation, including two clinically relevant C5 polymorphisms at R885.MethodsThe interaction of zilucoplan with C5, including for clinical C5 R885 variants, was investigated using surface plasmon resonance (SPR), hemolysis assays, and ELISA. The interference of C5b6 formation by zilucoplan was investigated by native gel analysis and hemolysis assay. The permeability of zilucoplan in a reconstituted basement membrane was assessed by the partition of zilucoplan on Matrigel-coated transwell chambers.ResultsZilucoplan specifically bound human complement C5 with high affinity, competitively inhibited the binding of C5 to C3b, and blocked C5 cleavage by C5 convertases and the assembly of the cytolytic membrane attack complex (MAC, or C5b9). Zilucoplan fully prevented the in vitro activation of C5 clinical variants at R885 that have been previously reported to respond poorly to eculizumab treatment. Zilucoplan was further demonstrated to interfere with the formation of C5b6 and inhibit red blood cell (RBC) hemolysis induced by plasmin-mediated non-canonical C5 activation. Zilucoplan demonstrated greater permeability than a monoclonal C5 antibody in a reconstituted basement membrane model, providing a rationale for the rapid onset of action of zilucoplan observed in clinical studies.ConclusionOur findings demonstrate that zilucoplan uses a dual mode of action to potently inhibit the activation of C5 and terminal complement pathway including wild-type and clinical R885 variants that do not respond to eculizumab treatment. These data may be relevant to the clinically demonstrated benefits of zilucoplan.

Crop yield prediction is essential for tasks like determining the optimal profile of crops to be planted, allocating government resources, effectively planning and preparing for aid distribution, making decisions about imports, and so on. Crop yield prediction using remote sensing data during the growing season is helpful to farm planning and management, which has received increasing attention. Information mining from multichannel geo-spatiotemporal data brings many benefits to crop yield prediction. However, most of the existing methods have not fully utilized the dimension reduction technology and the spatiotemporal feature of the data. In this paper, a new approach is proposed to predict the yield from multispatial images by using the dimension reduction method and a 3D convolutional neural network. In addition, regions with similar crop yields should have similar features learned by the network. Thus, metric learning and multitask learning are used to learn more discriminative features. We evaluated the proposed method on county-level soybean yield prediction in the United States, and the experimental results show the effectiveness of the proposed method. The proposed method provides new ideas for crop yield estimation and effectively improves the accuracy of crop yield estimation.

With the continuous advancement of the integration of the world’s digital economy, the economic development of various regions has become increasingly interdependent. However, the interregional interactions of digital industries have yet to be further elucidated. Here, based on the Multi-Regional Input-Output model, we take China, for example, and analyze the impact of digital industries on the economy from the industrial and regional. At the industry level, we discuss the economic spillover relationship between the digital industry and the three traditional industries, and analyzes the spatial and temporal differentiation in the inter-industry spillover status of China’s 30 provinces; at the regional level, we discuss the digital industrial economic spillover links and spillover states. The results show that: 1) The self-generating ability within the digital industry sector is the most significant factor influencing the industrial growth of China’s digital economy, followed by the interrelated effect between industry sectors, while the inter-industry feedback effect has a weak impact on the economic system. 2) There is a clear gap in the economic spillover capacity of the digital industry in 30 provinces, and the overall performance is decreasing in the eastern, central and western regions. 3) The intra-regional multiplier effect of digital industry shows a slowly decreasing trend, and the intra-regional digital economic spillover-feedback effect shows a growing trend. At the same time, the inter-regional digital economy interaction tends to decay in distance, indicating that regional accessibility has a significant impact on the inter-regional spillover effect.

Employees are important stakeholders in an organization. This paper aims to examine the effectiveness of limits on employee compensation in state-owned enterprises (SOEs), a policy for employees of state-owned enterprises issued by the China State-owned Assets Supervision and Administration Commission (SASAC) in 2010. Employing a difference-in-differences analysis for a sample of Chinese listed companies from 2007 to 2013, the results show that employee compensation restriction enhances the labor productivity of SOEs. This policy effect is mainly due to the contribution of compensation limits to the external fairness of employee compensation, and the findings remain unchanged after a series of robustness testing procedures. In addition, the employee compensation restriction policy significantly affects labor productivity improvement in monopolistic industries or mature SOEs.

The long-term prediction of the degradation of key computer parameters improves maintenance performance. Traditional prediction methods may suffer from cumulative errors in iterative prediction, which affect the model’s long-term prediction accuracy. Our network adopts curriculum learning and transfer learning methods, which can effectively solve this problem. The training network uses a dual-branch Siamese network. One branch intermixes the predicted and annotated data as input and uses curriculum learning to train. The other branch uses the original annotated data for training. To further align the hidden distributions of the two branches, the transfer learning method calculates the covariance matrices of the time series of the two branches by correlation alignment loss. A single branch is used in the test for prediction without increasing the inference computation. Compared with the current mainstream networks, our method can effectively improve the accuracy of long-term prediction with the improvements above.

Exploiting negative capacitance effects in organic thin-film transistors (OTFTs) is advantageous for enhancing device performance. Here, the authors report solution-processed sub-thermionic OTFTs and circuits with ferroelectric hafnium oxides that show ultra-low power and ultra-high gain.

Polymeric materials are emerging as an important component for organic electronics due to their combined features of soft materials and organic semiconductor/conductor. Different from solid-state bulk materials, polymers are synthesized and modulated with a facile method, with their controllable properties. Therefore, polymers have been recently used in Schottky junction devices that demonstrate features of enhanced performances, flexibility, and easy processing. Polymer-based Schottky junctions have drawn a lot of interest in several prototype applications like photodetectors, energy harvesters, and gas sensors. With this review, we summarize recent progress in synthesis and modulation of polymers and fabrication of polymer-based Schottky junction devices. Designs of applications of sensory devices like photodetectors, energy harvesters, and gas sensors are also presented.

The load resisting capacity of structural members will decrease when they are subjected to long-term sustained load. Such phenomenon is widely known as the duration of load effect, which is mainly caused by the damage accumulation in the material. The deterioration mechanism of the material is a typical stochastic process which is influenced by a large variety of parameters involving complex physical and chemical process. Although classical models have been proposed to evaluate the duration of load effect, it is nearly impossible to quantify the influence of various parameters and to achieve an accurate estimation. To optimize the combination of complexity and goodness-of-fit, a neural network model is proposed in this paper to evaluate the duration of load effect in wood structural members. Taking individual uncertainties into consideration, the proposed model treats the damage in wood as a Markov process and can estimate the residual strength distribution of the investigated wood structural members under long-term sustained load. The coefficient of determination reaches above 95% under sustained loading scenario, and it shows good adaptability across different wood properties. Moreover, the model can be adapted to continuously varied loading scenarios with a 98% coefficient of determination. This research aims to provide a useful and straightforward tool for accurately predicting the duration of load effect in wood structural members, and the proposed algorithm can be easily modified to deal with similar engineering problems for other construction materials.

Embedded computing platforms such as neural network accelerators deploying neural network models need to quantize the values into low-bit integers through quantization operations. However, most current embedded computing platforms with a fixed-point architecture do not directly support performing the quantization operation for the LSTM layer. Meanwhile, the influence of sequential input data for LSTM has not been taken into account by quantization algorithms. Aiming at these two technical bottlenecks, a new sequential-characteristics-based operators disassembly quantization method for LSTM layers is proposed. Specifically, the calculation process of the LSTM layer is split into multiple regular layers supported by the neural network accelerator. The quantization-parameter-generation process is designed as a sequential-characteristics-based combination strategy for sequential and diverse image groups. Therefore, LSTM is converted into multiple mature operators for single-layer quantization and deployed on the neural network accelerator. Comparison experiments with the state of the art show that the proposed quantization method has comparable or even better performance than the full-precision baseline in the field of character-/word-level language prediction and image classification applications. The proposed method has strong application potential in the subsequent addition of novel operators for future neural network accelerators.

Walnut (Juglans regia) is an important woody oil-bearing plant with high nutritional value. For better understanding of the underlying molecular mechanisms of its oil accumulation in the Qinghai Plateau, in this study we monitored walnut fruit development, and 15 cDNA libraries were constructed from walnut seed kernels collected at 72, 79, 93, 118 and 135 days after flowering (DAF). The candidate genes were identified using sequencing and expression analysis. The results showed that the oil content in the kernels increased dramatically in late July and reached the maximum value of 69% in mature seed. More than 90% of the oils were unsaturated fatty acids (UFAs) and linoleic acid (18:2) was the predominant UFA accumulated in mature seed. Differentially expressed genes (DEGs) in 15 KEGG pathways of lipid metabolism were detected. We identified 119 DEGs related to FA de novo biosynthesis (38 DEGs), FA elongation and desaturation (39 DEGs), triacylglycerol (TAG) assembly (24 DEGs), oil bodies (12 DEGs), and transcription factors (TFs, 6 DEGs). The abundantly expressed oleosins, caleosins and steroleosins may be important for timely energy reserve in oil bodies. Weighted gene coexpression network analysis (WGCNA) showed that AP2/ERF and bHLH were the key TFs, and were co-expressed with ACC1, α-CT, BCCP, MAT, KASII, LACS, FATA, and PDCT. Our transcriptome data will enrich public databases and provide new insights into functional genes related to the seed kernel lipid metabolism and oil accumulation in J. regia.

Object tracking based on deep learning is a hot topic in computer vision with many applications. Due to high computation and memory costs, it is difficult to deploy convolutional neural networks (CNNs) for object tracking on embedded systems with limited hardware resources. This paper uses the Siamese network to construct the backbone of our tracker. The convolution layers used to extract features often have the highest costs, so more improvements should be focused on them to make the tracking more efficient. In this paper, the standard convolution is optimized by the separable convolution, which mainly includes a depthwise convolution and a pointwise convolution. To further reduce the calculation, filters in the depthwise convolution layer are pruned with filters variance. As there are different weight distributions in convolution layers, the filter pruning is guided by a hyper-parameter designed. With the improvements, the number of parameters is decreased to 13% of the original network and the computation is reduced to 23%. On the NVIDIA Jetson TX2, the tracking speed increased to 3.65 times on the CPU and 2.08 times on the GPU, without significant degradation of tracking performance in VOT benchmark.

Yin-mountain Hilly Area is one of the ideal regions for potato (Solanum tuberosum) production in China. However, potato yield is severely limited as a result of rain-fed crop mode due to water deficiency, as well as an inadequate farming practices. In this study, yield gaps were determined by using attainable yield (Ya) as a benchmark under optimized management practices, i.e., micro-ridge and side planting with plastic-mulching (MS), and flat planting with plastic-mulching (PM). The yields under MS and PM modes are defined as Ya1 and Ya2, respectively. Under the same field with MS and PM modes but different densities and fertilizer usages and so on, it was defined as simulated farmers' practices. The yield of simulated farmers' practices (Yf1) reached 57.3 and 69.6% of Ya1 and Ya2, respectively, while the average yield of 298 randomly surveyed farmers (Yf2) reached only 37.0 and 47.8% of Ya1 and Ya2 for rain-fed potato, respectively. The gaps of water use efficiency exhibited similar pattern. Further analysis shows that improper measures in rainwater conservation and accumulation, and other management practices contributed to 18.5, 18.2, and 42.6% of yield gap between Ya1 and Yf2. Improper nutrition management, including overuse of nitrogen and the deficiency of phosphorus and potassium supplication, was one of the important reasons of yield gap. The results indicate the possibilities of increasing rain-fed potato yields by optimized water and fertilizer managements in the Yin-mountain Hilly Area.

MXenes, a kind of two-dimensional material of early transition metal carbides and carbonitrides, have emerged as a unique class of layered-structured metallic materials with attractive features, as good conductivity comparable to metals, enhanced ionic conductivity, hydrophilic property derived from their hydroxyl or oxygen-terminated surfaces, and mechanical flexibility. With tunable etching methods, the morphology of MXenes can be effectively controlled to form nanoparticles, single layer, or multi-layer nanosheets, which exhibit large specific surface areas and is favorable for enhancing the sensing performance of MXenes based sensors. Moreover, MXenes are available to form composites with other materials facilely. With structure design, MXenes or its composite show enhanced mechanical flexibility and stretchability, which enabled its wide application in the fields of wearable sensors, energy storage, and electromagnetic shielding. In this review, recent progress in MXenes is summarized, focusing on its application in wearable sensors including pressure/strain sensing, biochemical sensing, temperature, and gas sensing. Furthermore, the main challenges and future research are also discussed.

Local features extraction is a crucial technology for image matching navigation of an unmanned aerial vehicle (UAV), where it aims to accurately and robustly match a real-time image and a geo-referenced image to obtain the position update information of the UAV. However, it is a challenging task due to the inconsistent image capture conditions, which will lead to extreme appearance changes, especially the different imaging principle between an infrared image and RGB image. In addition, the sparsity and labeling complexity of existing public datasets hinder the development of learning-based methods in this research area. This paper proposes a novel learning local features extraction method, which uses local features extracted by deep neural network to find the correspondence features on the satellite RGB reference image and real-time infrared image. First, we propose a single convolution neural network that simultaneously extracts dense local features and their corresponding descriptors. This network combines the advantages of a high repeatability local feature detector and high reliability local feature descriptors to match the reference image and real-time image with extreme appearance changes. Second, to make full use of the sparse dataset, an iterative training scheme is proposed to automatically generate the high-quality corresponding features for algorithm training. During the scheme, the dense correspondences are automatically extracted, and the geometric constraints are added to continuously improve the quality of them. With these improvements, the proposed method achieves state-of-the-art performance for infrared aerial (UAV captured) image and satellite reference image, which shows 4–6% performance improvements in precision, recall, and F1-score, compared to the other methods. Moreover, the applied experiment results show its potential and effectiveness on localization for UAVs navigation and trajectory reconstruction application.

Altitude estimation is one of the fundamental tasks of unmanned aerial vehicle (UAV) automatic navigation, where it aims to accurately and robustly estimate the relative altitude between the UAV and specific areas. However, most methods rely on auxiliary signal reception or expensive equipment, which are not always available, or applicable owing to signal interference, cost or power-consuming limitations in real application scenarios. In addition, fixed-wing UAVs have more complex kinematic models than vertical take-off and landing UAVs. Therefore, an altitude estimation method which can be robustly applied in a GPS denied environment for fixed-wing UAVs must be considered. In this paper, we present a method for high-precision altitude estimation that combines the vision information from a monocular camera and poses information from the inertial measurement unit (IMU) through a novel end-to-end deep neural network architecture. Our method has numerous advantages over existing approaches. First, we utilize the visual-inertial information and physics-based reasoning to build an ideal altitude model that provides general applicability and data efficiency for neural network learning. A further advantage is that we have designed a novel feature fusion module to simplify the tedious manual calibration and synchronization of the camera and IMU, which are required for the standard visual or visual-inertial methods to obtain the data association for altitude estimation modeling. Finally, the proposed method was evaluated, and validated using real flight data obtained during a fixed-wing UAV landing phase. The results show the average estimation error of our method is less than 3% of the actual altitude, which vastly improves the altitude estimation accuracy compared to other visual and visual-inertial based methods.

Worldwide, a huge amount of solid fermented waste is produced every year during mushroom production. The rational utilization of spent mushroom substrate (SMS) is conducive to environmental protection and the sustainable development of agriculture. The aims of this work were to analyze the effects of the SMS mulching thickness on the soil environment, weed suppression, leaf nutrients, and nut traits in a hazelnut plantation and provide a theoretical basis for the scientific and reasonable utilization of SMS. An SMS mulching experiment with four mulching thickness treatments (MT1: 5 cm, MT2: 10 cm, MT3: 15 cm, and CK: 0 cm) was conducted in a semi-arid area of North China in 2019. The soil properties, weed control efficiency, leaf nutrient content, and nut yield and quality characteristics of these treatments were all measured and analyzed. The results showed that (1) the average soil temperatures of MT2 and MT3 at different soil depths were 0.73–1.78 °C and 1.18–2.50 °C lower than when no mulch was used in warm weather and 0.67–1.05 °C and 0.99–1.56 °C higher than when no mulch was used in cold weather. The average soil moisture content of MT1, MT2, and MT3 at different soil depths was 6.27–8.13%, 10.66–17.23%, and 7.26–12.81% higher than that of CK, respectively. There were no significant differences in the soil bulk density or porosity among the four treatments. The average soil nutrient contents (soil organic matter and available N, P, and K) of MT1, MT2, and MT3 were 7.05–15.23%, 14.90–38.93%, and 17.42–40.11% higher than those of CK, respectively. Soil enzymatic activities of these treatments followed the order MT3 ≈ MT2 > MT1 > CK. (2) MT2 and MT3 both had high levels of weed control efficiency: 55.66–92.93% and 70.43–97.90%, respectively. (3) The leaf N, P, K, and SPAD increased significantly under MT2 and MT3 by about 10% compared with the CK. (4) MT2 effectively increased the nut size, nut and kernel mass, and crude fat content of the nut (CFC). In general, the short-term field experiment clearly showed that SMS is a superior mulching material for improving the soil environment and plant growth, and 10 cm thickness is suitable for mulching practice.

Near-field communication is a new kind of low-cost wireless communication technology developed in recent years, which brings great convenience to daily life activities such as medical care, food quality detection, and commerce. The integration of near-field communication devices and sensors exhibits great potential for these real-world applications by endowing sensors with new features of powerless and wireless signal transferring and conferring near field communication device with sensing function. In this review, we summarize recent progress in near field communication sensors, including the development of materials and device design and their applications in wearable personal healthcare devices. The opportunities and challenges in near-field communication sensors are discussed in the end.

Ammonia (NH3) is a harmful gas contaminant that is part of the nitrogen cycle in our daily lives. Therefore, highly sensitive ammonia sensors are important for environmental protection and human health. However, it is difficult to detect low concentrations of ammonia (≤50 ppm) using conventional means at room temperature. Tin monoxide (SnO), a member of IV–VI metal monoxides, has attracted much attention due to its low cost, environmental-friendly nature, and higher stability compared with other non-oxide ammonia sensing material like alkaline metal or polymer, which made this material an ideal alternative for ammonia sensor applications. In this work, we fabricated high-sensitive ammonia sensors based on self-assembly SnO nanoshells via a solution method and annealing under 300 °C for 1 h. The as fabricated sensors exhibited the response of 313%, 874%, 2757%, 3116%, and 3757% (∆G/G) under ammonia concentration of 5 ppm, 20 ppm, 50 ppm, 100 ppm, and 200 ppm, respectively. The structure of the nanoshells, which have curved shells that provide shelters for the core and also possess a large surface area, is able to absorb more ammonia molecules, leading to further improvements in the sensitivity. Further, the SnO nanoshells have higher oxygen vacancy densities compared with other metal oxide ammonia sensing materials, enabling it to have higher performance. Additionally, the selectivity of ammonia sensors is also outstanding. We hope this work will provide a reference for the study of similar structures and applications of IV–VI metal monoxides in the gas sensor field.

When a satellite performs complex tasks such as discarding a payload or capturing a non-cooperative target, it will encounter sudden changes in the attitude and mass parameters, causing unstable flying and rolling of the satellite. In such circumstances, the change of the movement and mass characteristics are unpredictable. Thus, the traditional attitude control methods are unable to stabilize the satellite since they are dependent on the mass parameters of the controlled object. In this paper, we proposed a reinforcement learning method to re-stabilize the attitude of a satellite under such circumstances. Specifically, we discretize the continuous control torque, and build a neural network model that can output the discretized control torque to control the satellite. A dynamics simulation environment of the satellite is built, and the deep Q Network algorithm is then performed to train the neural network in this simulation environment. The reward of the training is the stabilization of the satellite. Simulation experiments illustrate that, with the iteration of training progresses, the neural network model gradually learned to re-stabilize the attitude of a satellite after unknown disturbance. As a contrast, the traditional PD (Proportion Differential) controller was unable to re-stabilize the satellite due to its dependence on the mass parameters. The proposed method adopts self-learning to control satellite attitudes, shows considerable intelligence and certain universality, and has a strong application potential for future intelligent control of satellites performing complex space tasks.

Semantic-level land-use scene classification is a challenging problem, in which deep learning methods, e.g., convolutional neural networks (CNNs), have shown remarkable capacity. However, a lack of sufficient labeled images has proved a hindrance to increasing the land-use scene classification accuracy of CNNs. Aiming at this problem, this paper proposes a CNN pre-training method under the guidance of a human visual attention mechanism. Specifically, a computational visual attention model is used to automatically extract salient regions in unlabeled images. Then, sparse filters are adopted to learn features from these salient regions, with the learnt parameters used to initialize the convolutional layers of the CNN. Finally, the CNN is further fine-tuned on labeled images. Experiments are performed on the UCMerced and AID datasets, which show that when combined with a demonstrative CNN, our method can achieve 2.24% higher accuracy than a plain CNN and can obtain an overall accuracy of 92.43% when combined with AlexNet. The results indicate that the proposed method can effectively improve CNN performance using easy-to-access unlabeled images and thus will enhance the performance of land-use scene classification especially when a large-scale labeled dataset is unavailable.

The engulfment of dying cells is a specialized form of phagocytosis that is extremely conserved across evolution. In the worm, it is genetically controlled by two parallel pathways, which are only partially reconstituted in mammals. We focused on the recapitulation of the CED-1 defined pathway in mammalian systems. We first explored and validated MEGF10, a novel receptor bearing striking structural similarities to CED-1, as a bona fide functional ortholog in mammals and hence progressed toward the analysis of molecular interactions along the corresponding pathway. We ascertained that, in a system of forced expression by transfection, MEGF10 function can be modulated by the ATP binding cassette transporter ABCA1, ortholog to CED-7. Indeed, the coexpression of either a functional or a mutant ABCA1 exerted a transdominant positive or negative modulation on the MEGF10-dependent engulfment. The combined use of biochemical and biophysical approaches indicated that this functional cooperation relies on the alternate association of these receptors with a common partner, endogenously expressed in our cell system. We provide the first working model structuring in mammals the CED-1 dependent pathway.

The hydrazone functional group, when coupled with a pyridyl substituent, offers a unique class of widely tunable photoswitches, whose E-to-Z photoisomerization equilibria can be controlled through intramolecular hydrogen bonding between the N–H hydrazone donor and the pyridyl acceptor. However, little is known about the photoisomerization mechanism in this class of compounds. To address this issue, we report a pyridine-appended iminoguanidinium photoswitch that is functionally related to acylhydrazones and provides insight into the photoisomerization processes between the E and Z configurations. The E-to-Z photoisomerization of the E-2-pyridyl-iminoguanidinium cation (2PyMIG) in DMSO, prepared as the bromide salt, was quantified by ¹H NMR, and probed in real time with ultrafast laser spectroscopy. The photoisomerization process occurs on a picosecond timescale, resulting in low fluorescence yields. The multiconfigurational reaction path found with the growing string method features a small barrier (4.3 or 6.5 kcal mol−1) that the E isomer in the π–π* state must overcome to reach the minimum energy conical intersection (MECI) connecting the E and Z isomers of 2PyMIG. While two possible pathways exist depending on the orientation of the pyridine ring, both exhibit the same qualitative features along the path and at their MECIs, involving simultaneous changes in the CCNN and CNNC dihedral angles. Furthermore, the ground state barrier for pyridine ring rotation is readily accessible, thus a low barrier pathway to the experimentally observed Z isomer exists for both MECIs leading to a transition from the E isomer to photoproduct. Combining multiconfigurational reaction path calculations using growing string method with time-resolved fluorescence spectroscopy provided crucial insights into the photoisomerization process of 2PyMIG, resulting in both the computational and experimental results pointing to rapid photoisomerization via a surface crossing between the singlet π–π* and the ground states. [ABSTRACT FROM AUTHOR]

The quantization algorithm compresses the original network by reducing the numerical bit width of the model, which improves the computation speed. Because different layers have different redundancy and sensitivity to data bit width. Reducing the data bit width will result in a loss of accuracy. Therefore, it is difficult to determine the optimal bit width for different parts of the network with guaranteed accuracy. Mixed precision quantization can effectively reduce the amount of computation while keeping the model accuracy basically unchanged. In this paper, a hardware-aware mixed precision quantization strategy optimal assignment algorithm adapted to low bit width is proposed, and reinforcement learning is used to automatically predict the mixed precision that meets the constraints of hardware resources. In the state-space design, the standard deviation of weights is used to measure the distribution difference of data, the execution speed feedback of simulated neural network accelerator inference is used as the environment to limit the action space of the agent, and the accuracy of the quantization model after retraining is used as the reward function to guide the agent to carry out deep reinforcement learning training. The experimental results show that the proposed method obtains a suitable model layer-by-layer quantization strategy under the condition that the computational resources are satisfied, and themodel accuracy is effectively improved. The proposed method has strong intelligence and certain universality and has strong application potential in the field of mixed precision quantization and embedded neural network model deployment. [ABSTRACT FROM AUTHOR]