In this paper, the optimization of quantizer’s segment threshold is done. The quantizer is designed on the basis of approximative spline functions. Coefficients on which we form approximative spline functions are cal...In this paper, the optimization of quantizer’s segment threshold is done. The quantizer is designed on the basis of approximative spline functions. Coefficients on which we form approximative spline functions are calculated by minimization mean square error (MSE). For coefficients determined in this way, spline functions by which optimal compressor function is approximated are obtained. For the quantizer designed on the basis of approximative spline functions, segment threshold is numerically determined depending on maximal value of the signal to quantization noise ratio (SQNR). Thus, quantizer with optimized segment threshold is achieved. It is shown that by quantizer model designed in this way and proposed in this paper, the SQNR that is very close to SQNR of nonlinear optimal companding quantizer is achieved.展开更多
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 d...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 databit width. Reducing the data bit width will result in a loss of accuracy. Therefore, it is difficult to determinethe optimal bit width for different parts of the network with guaranteed accuracy. Mixed precision quantizationcan effectively reduce the amount of computation while keeping the model accuracy basically unchanged. In thispaper, a hardware-aware mixed precision quantization strategy optimal assignment algorithm adapted to low bitwidth is proposed, and reinforcement learning is used to automatically predict the mixed precision that meets theconstraints of hardware resources. In the state-space design, the standard deviation of weights is used to measurethe distribution difference of data, the execution speed feedback of simulated neural network accelerator inferenceis used as the environment to limit the action space of the agent, and the accuracy of the quantization model afterretraining is used as the reward function to guide the agent to carry out deep reinforcement learning training. Theexperimental results show that the proposed method obtains a suitable model layer-by-layer quantization strategyunder 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 thefield of mixed precision quantization and embedded neural network model deployment.展开更多
The demand for adopting neural networks in resource-constrained embedded devices is continuously increasing.Quantization is one of the most promising solutions to reduce computational cost and memory storage on embedd...The demand for adopting neural networks in resource-constrained embedded devices is continuously increasing.Quantization is one of the most promising solutions to reduce computational cost and memory storage on embedded devices.In order to reduce the complexity and overhead of deploying neural networks on Integeronly hardware,most current quantization methods use a symmetric quantization mapping strategy to quantize a floating-point neural network into an integer network.However,although symmetric quantization has the advantage of easier implementation,it is sub-optimal for cases where the range could be skewed and not symmetric.This often comes at the cost of lower accuracy.This paper proposed an activation redistribution-based hybrid asymmetric quantizationmethod for neural networks.The proposedmethod takes data distribution into consideration and can resolve the contradiction between the quantization accuracy and the ease of implementation,balance the trade-off between clipping range and quantization resolution,and thus improve the accuracy of the quantized neural network.The experimental results indicate that the accuracy of the proposed method is 2.02%and 5.52%higher than the traditional symmetric quantization method for classification and detection tasks,respectively.The proposed method paves the way for computationally intensive neural network models to be deployed on devices with limited computing resources.Codes will be available on https://github.com/ycjcy/Hybrid-Asymmetric-Quantization.展开更多
The recently developed magic-intensity trapping technique of neutral atoms efficiently mitigates the detrimental effect of light shifts on atomic qubits and substantially enhances the coherence time. This technique re...The recently developed magic-intensity trapping technique of neutral atoms efficiently mitigates the detrimental effect of light shifts on atomic qubits and substantially enhances the coherence time. This technique relies on applying a bias magnetic field precisely parallel to the wave vector of a circularly polarized trapping laser field. However, due to the presence of the vector light shift experienced by the trapped atoms, it is challenging to precisely define a parallel magnetic field, especially at a low bias magnetic field strength, for the magic-intensity trapping of85Rb qubits. In this work, we present a method to calibrate the angle between the bias magnetic field and the trapping laser field with the compensating magnetic fields in the other two directions orthogonal to the bias magnetic field direction. Experimentally, with a constantdepth trap and a fixed bias magnetic field, we measure the respective resonant frequencies of the atomic qubits in a linearly polarized trap and a circularly polarized one via the conventional microwave Rabi spectra with different compensating magnetic fields and obtain the corresponding total magnetic fields via the respective resonant frequencies using the Breit–Rabi formula. With known total magnetic fields, the angle is a function of the other two compensating magnetic fields.Finally, the projection value of the angle on either of the directions orthogonal to the bias magnetic field direction can be reduced to 0(4)° by applying specific compensating magnetic fields. The measurement error is mainly attributed to the fluctuation of atomic temperature. Moreover, it also demonstrates that, even for a small angle, the effect is strong enough to cause large decoherence of Rabi oscillation in a magic-intensity trap. Although the compensation method demonstrated here is explored for the magic-intensity trapping technique, it can be applied to a variety of similar precision measurements with trapped neutral atoms.展开更多
The nanoscale confinement is of great important for the industrial applications of molecular sieve,desalination,and also essential in bio-logical transport systems.Massive efforts have been devoted to the influence of...The nanoscale confinement is of great important for the industrial applications of molecular sieve,desalination,and also essential in bio-logical transport systems.Massive efforts have been devoted to the influence of restricted spaces on the properties of confined fluids.However,the situation of channel-wall is crucial but attracts less attention and remains unknown.To fundamentally understand the mechanism of channel-walls in nanoconfinement,we investigated the interaction between the counter-force of the liquid and interlamellar spacing of nanochannel walls by considering the effect of both spatial confinement and surface wettability.The results reveal that the nanochannel stables at only a few discrete spacing states when its confinement is within 1.4 nm.The quantized interlayer spacing is attributed to water molecules becoming laminated structures,and the stable states are corresponding to the monolayer,bilayer and trilayer water configurations,respectively.The results can potentially help to understand the characterized interlayers spacing of graphene oxide membrane in water.Our findings are hold great promise in design of ion filtration membrane and artificial water/ion channels.展开更多
This paper is concerned with distributed Nash equi librium seeking strategies under quantized communication. In the proposed seeking strategy, a projection operator is synthesized with a gradient search method to achi...This paper is concerned with distributed Nash equi librium seeking strategies under quantized communication. In the proposed seeking strategy, a projection operator is synthesized with a gradient search method to achieve the optimization o players' objective functions while restricting their actions within required non-empty, convex and compact domains. In addition, a leader-following consensus protocol, in which quantized informa tion flows are utilized, is employed for information sharing among players. More specifically, logarithmic quantizers and uniform quantizers are investigated under both undirected and connected communication graphs and strongly connected digraphs, respec tively. Through Lyapunov stability analysis, it is shown that play ers' actions can be steered to a neighborhood of the Nash equilib rium with logarithmic and uniform quantizers, and the quanti fied convergence error depends on the parameter of the quan tizer for both undirected and directed cases. A numerical exam ple is given to verify the theoretical results.展开更多
Quantized training has been proven to be a prominent method to achieve deep neural network training under limited computational resources.It uses low bit-width arithmetics with a proper scaling factor to achieve negli...Quantized training has been proven to be a prominent method to achieve deep neural network training under limited computational resources.It uses low bit-width arithmetics with a proper scaling factor to achieve negligible accuracy loss.Cambricon-Q is the ASIC design proposed to efficiently support quantized training,and achieves significant performance improvement.However,there are still two caveats in the design.First,Cambricon-Q with different hardware specifications may lead to different numerical errors,resulting in non-reproducible behaviors which may become a major concern in critical applications.Second,Cambricon-Q cannot leverage data sparsity,where considerable cycles could still be squeezed out.To address the caveats,the acceleration core of Cambricon-Q is redesigned to support fine-grained irregular data processing.The new design not only enables acceleration on sparse data,but also enables performing local dynamic quantization by contiguous value ranges(which is hardware independent),instead of contiguous addresses(which is dependent on hardware factors).Experimental results show that the accuracy loss of the method still keeps negligible,and the accelerator achieves 1.61×performance improvement over Cambricon-Q,with about 10%energy increase.展开更多
The singularity at distance r → 0 at the center of a spherically symmetric non-rotating, uncharged mass of radius R, is considered here. Under inverse square law force, the Schwarzschild metric, needs to be modified,...The singularity at distance r → 0 at the center of a spherically symmetric non-rotating, uncharged mass of radius R, is considered here. Under inverse square law force, the Schwarzschild metric, needs to be modified, to include Newton’s Shell Theorem (NST). By including NST for r, both Schwarzschild singularity at r = 2GM/c2 and at r → 0 singularities are removed from the metric. Near R → 0, the question of maximal density is considered based on Schwarzschild’s modified metric, and compared to the quantum limit of maximal mass density put by Planck’s quantum-based universal units. It is asserted, that General relativity, when combined with Planck’s universal units, inevitably leads to quantization of gravity.展开更多
基金Serbian Ministry of Education and Science through Mathematical Institute of Serbian Academy of Sciences and Arts(Project III44006)Serbian Ministry of Education and Science(Project TR32035)
文摘In this paper, the optimization of quantizer’s segment threshold is done. The quantizer is designed on the basis of approximative spline functions. Coefficients on which we form approximative spline functions are calculated by minimization mean square error (MSE). For coefficients determined in this way, spline functions by which optimal compressor function is approximated are obtained. For the quantizer designed on the basis of approximative spline functions, segment threshold is numerically determined depending on maximal value of the signal to quantization noise ratio (SQNR). Thus, quantizer with optimized segment threshold is achieved. It is shown that by quantizer model designed in this way and proposed in this paper, the SQNR that is very close to SQNR of nonlinear optimal companding quantizer is achieved.
文摘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 databit width. Reducing the data bit width will result in a loss of accuracy. Therefore, it is difficult to determinethe optimal bit width for different parts of the network with guaranteed accuracy. Mixed precision quantizationcan effectively reduce the amount of computation while keeping the model accuracy basically unchanged. In thispaper, a hardware-aware mixed precision quantization strategy optimal assignment algorithm adapted to low bitwidth is proposed, and reinforcement learning is used to automatically predict the mixed precision that meets theconstraints of hardware resources. In the state-space design, the standard deviation of weights is used to measurethe distribution difference of data, the execution speed feedback of simulated neural network accelerator inferenceis used as the environment to limit the action space of the agent, and the accuracy of the quantization model afterretraining is used as the reward function to guide the agent to carry out deep reinforcement learning training. Theexperimental results show that the proposed method obtains a suitable model layer-by-layer quantization strategyunder 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 thefield of mixed precision quantization and embedded neural network model deployment.
基金The Qian Xuesen Youth Innovation Foundation from China Aerospace Science and Technology Corporation(Grant Number 2022JY51).
文摘The demand for adopting neural networks in resource-constrained embedded devices is continuously increasing.Quantization is one of the most promising solutions to reduce computational cost and memory storage on embedded devices.In order to reduce the complexity and overhead of deploying neural networks on Integeronly hardware,most current quantization methods use a symmetric quantization mapping strategy to quantize a floating-point neural network into an integer network.However,although symmetric quantization has the advantage of easier implementation,it is sub-optimal for cases where the range could be skewed and not symmetric.This often comes at the cost of lower accuracy.This paper proposed an activation redistribution-based hybrid asymmetric quantizationmethod for neural networks.The proposedmethod takes data distribution into consideration and can resolve the contradiction between the quantization accuracy and the ease of implementation,balance the trade-off between clipping range and quantization resolution,and thus improve the accuracy of the quantized neural network.The experimental results indicate that the accuracy of the proposed method is 2.02%and 5.52%higher than the traditional symmetric quantization method for classification and detection tasks,respectively.The proposed method paves the way for computationally intensive neural network models to be deployed on devices with limited computing resources.Codes will be available on https://github.com/ycjcy/Hybrid-Asymmetric-Quantization.
基金Project supported by the National Natural Science Foundation of China(Grant Nos.12104414,12122412,12104464,and 12104413)the China Postdoctoral Science Foundation(Grant No.2021M702955).
文摘The recently developed magic-intensity trapping technique of neutral atoms efficiently mitigates the detrimental effect of light shifts on atomic qubits and substantially enhances the coherence time. This technique relies on applying a bias magnetic field precisely parallel to the wave vector of a circularly polarized trapping laser field. However, due to the presence of the vector light shift experienced by the trapped atoms, it is challenging to precisely define a parallel magnetic field, especially at a low bias magnetic field strength, for the magic-intensity trapping of85Rb qubits. In this work, we present a method to calibrate the angle between the bias magnetic field and the trapping laser field with the compensating magnetic fields in the other two directions orthogonal to the bias magnetic field direction. Experimentally, with a constantdepth trap and a fixed bias magnetic field, we measure the respective resonant frequencies of the atomic qubits in a linearly polarized trap and a circularly polarized one via the conventional microwave Rabi spectra with different compensating magnetic fields and obtain the corresponding total magnetic fields via the respective resonant frequencies using the Breit–Rabi formula. With known total magnetic fields, the angle is a function of the other two compensating magnetic fields.Finally, the projection value of the angle on either of the directions orthogonal to the bias magnetic field direction can be reduced to 0(4)° by applying specific compensating magnetic fields. The measurement error is mainly attributed to the fluctuation of atomic temperature. Moreover, it also demonstrates that, even for a small angle, the effect is strong enough to cause large decoherence of Rabi oscillation in a magic-intensity trap. Although the compensation method demonstrated here is explored for the magic-intensity trapping technique, it can be applied to a variety of similar precision measurements with trapped neutral atoms.
基金support from the National Natural Science Foundation of China(Grant Nos.12372327,12372109,11972171)National Key R&D Program of China(Grant No.2023YFB4605101).
文摘The nanoscale confinement is of great important for the industrial applications of molecular sieve,desalination,and also essential in bio-logical transport systems.Massive efforts have been devoted to the influence of restricted spaces on the properties of confined fluids.However,the situation of channel-wall is crucial but attracts less attention and remains unknown.To fundamentally understand the mechanism of channel-walls in nanoconfinement,we investigated the interaction between the counter-force of the liquid and interlamellar spacing of nanochannel walls by considering the effect of both spatial confinement and surface wettability.The results reveal that the nanochannel stables at only a few discrete spacing states when its confinement is within 1.4 nm.The quantized interlayer spacing is attributed to water molecules becoming laminated structures,and the stable states are corresponding to the monolayer,bilayer and trilayer water configurations,respectively.The results can potentially help to understand the characterized interlayers spacing of graphene oxide membrane in water.Our findings are hold great promise in design of ion filtration membrane and artificial water/ion channels.
基金supported by the National Natural Science Foundation of China (NSFC)(62222308, 62173181, 62073171, 62221004)the Natural Science Foundation of Jiangsu Province (BK20200744, BK20220139)+3 种基金Jiangsu Specially-Appointed Professor (RK043STP19001)the Young Elite Scientists Sponsorship Program by CAST (2021QNRC001)1311 Talent Plan of Nanjing University of Posts and Telecommunicationsthe Fundamental Research Funds for the Central Universities (30920032203)。
文摘This paper is concerned with distributed Nash equi librium seeking strategies under quantized communication. In the proposed seeking strategy, a projection operator is synthesized with a gradient search method to achieve the optimization o players' objective functions while restricting their actions within required non-empty, convex and compact domains. In addition, a leader-following consensus protocol, in which quantized informa tion flows are utilized, is employed for information sharing among players. More specifically, logarithmic quantizers and uniform quantizers are investigated under both undirected and connected communication graphs and strongly connected digraphs, respec tively. Through Lyapunov stability analysis, it is shown that play ers' actions can be steered to a neighborhood of the Nash equilib rium with logarithmic and uniform quantizers, and the quanti fied convergence error depends on the parameter of the quan tizer for both undirected and directed cases. A numerical exam ple is given to verify the theoretical results.
基金the National Key Research and Devecopment Program of China(No.2022YFB4501601)the National Natural Science Foundation of China(No.62102398,U20A20227,62222214,62002338,U22A2028,U19B2019)+1 种基金the Chinese Academy of Sciences Project for Young Scientists in Basic Research(YSBR-029)Youth Innovation Promotion Association Chinese Academy of Sciences。
文摘Quantized training has been proven to be a prominent method to achieve deep neural network training under limited computational resources.It uses low bit-width arithmetics with a proper scaling factor to achieve negligible accuracy loss.Cambricon-Q is the ASIC design proposed to efficiently support quantized training,and achieves significant performance improvement.However,there are still two caveats in the design.First,Cambricon-Q with different hardware specifications may lead to different numerical errors,resulting in non-reproducible behaviors which may become a major concern in critical applications.Second,Cambricon-Q cannot leverage data sparsity,where considerable cycles could still be squeezed out.To address the caveats,the acceleration core of Cambricon-Q is redesigned to support fine-grained irregular data processing.The new design not only enables acceleration on sparse data,but also enables performing local dynamic quantization by contiguous value ranges(which is hardware independent),instead of contiguous addresses(which is dependent on hardware factors).Experimental results show that the accuracy loss of the method still keeps negligible,and the accelerator achieves 1.61×performance improvement over Cambricon-Q,with about 10%energy increase.
文摘The singularity at distance r → 0 at the center of a spherically symmetric non-rotating, uncharged mass of radius R, is considered here. Under inverse square law force, the Schwarzschild metric, needs to be modified, to include Newton’s Shell Theorem (NST). By including NST for r, both Schwarzschild singularity at r = 2GM/c2 and at r → 0 singularities are removed from the metric. Near R → 0, the question of maximal density is considered based on Schwarzschild’s modified metric, and compared to the quantum limit of maximal mass density put by Planck’s quantum-based universal units. It is asserted, that General relativity, when combined with Planck’s universal units, inevitably leads to quantization of gravity.