Hessian-based mixed-precision quantization with transition aware training for neural networks.

Energy-efficient processors are crucial for accelerating deep neural networks (DNNs) on edge devices with limited battery capacity. To reduce energy consumption, time-domain computing-in-memory (TD-CIM) is a splendid architecture, which consumes low computation and memory access energy due to low toggle rate of time-based signals and less data movements, respectively. When deploying DNNs in TD-CIMs, quantization is required, which has two types: uniform quantization (UQ) and nonuniform quantization (NUQ). To reach the same accuracy for one DNN, NUQ achieves smaller model size than UQ. Due to varying weight distributions across layers, mixed-precision quantization can further reduce model size, without degrading accuracy. However, previous TD-CIMs are inefficient for mixed-precision NUQ-DNNs due to their adopted bit-serial convolution increasing computation amount significantly. To address that, we propose a unique-weight convolution to accelerate mixed-precision NUQ-DNNs by a special kernel decomposition, reducing computation count remarkably. Based on that, we design a TD-CIM-based processor, TIMAQ, with three architectural techniques: 1) bit-cross-flipping-based kernel decomposer to reduce memory accesses and operations of decomposing kernels; 2) dual-mode-complementary predictor to remove redundant computations; and 3) activation-weight-adaptive pulse quantizer to decrease pulse quantization energy and error. Fabricated in 28-nm CMOS technology and tested on 1–8-b NUQ-DNNs, TIMAQ achieves 2.4–152.7-TOPS/W peak energy efficiency.

Extensive research on deep neural networks for LiDAR point clouds has contributed inexhaustible momentum to the development of computer 3D vision applications. However, storage and energy consumption have always been a challenge for deploying these deep models on mobile devices. Quantization provides a feasible route, of which current primary research is focused on uniform bit-width quantization without considering different layers or filters’ sensitivity to different bit-widths. This article proposes a novel hybrid compression method based on relaxed mixed-precision quantization, relaxed weights pruning, and knowledge distillation to overcome the limitations of uniform quantization illustrated above, while further improving model accuracy and reducing model memory consumption. It employs a differentiable searching method to search for the optimal bit allocation and weight sparsity, while conducting feature distillation, accordingly considering the feature degradation by pooling operation in point cloud deep models. The proposed method combines three compression techniques to balance the trade-off between compression accuracy and model size. Pruning alleviates the increasing memory consumption problem caused by mixed-precision quantization, while distillation improves compression accuracy without increasing model size. The experiments validate that the proposed method outperforms state-of-the-art typical uniform quantization methods in terms of accuracy with an acceptable and relatively competitive compression performance.

While Deep Neural Networks (DNNs) quantization leads to a significant reduction in computational and storage costs, it reduces model capacity and therefore, usually leads to an accuracy drop. One of the possible ways to overcome this issue is to use different quantization bit-widths for different layers. The main challenge of the mixed-precision approach is to define the bit-widths for each layer, while staying under memory and latency requirements. Motivated by this challenge, we introduce a novel technique for explicit complexity control of DNNs quantized to mixed-precision, which uses smooth optimization on the surface containing neural networks of constant size. Furthermore, we introduce a family of smooth quantization regularizers, which can be used jointly with our complexity control method for both post-training mixed-precision quantization and quantization-aware training. Our approach can be applied to any neural network architecture. Experiments show that the proposed techniques reach state-of-the-art results. KeywordsNeural network quantizationMixed-precision quantizationRegularization for quantization

Mixed-precision quantization with adaptive bitwidth allocation for neural network has achieved higher compression rate and accuracy in classification task. However, it has not been well explored for object detection networks. In this paper, we propose a novel mixed-precision quantization scheme with dynamical Hessian matrix for object detection networks. We iteratively select a layer with the lowest sensitivity based on the Hessian matrix and downgrade its precision to reach the required compression ratio. The L-BFGS algorithm is utilized for updating the Hessian matrix in each quantization iteration. Moreover, we specifically design the loss function for objection detection networks by jointly considering the quantization effects on classification and regression loss. Experimental results on RetinaNet and Faster R-CNN show that the proposed DHMQ achieves state-of-the-art performance for quantized object detec-tors.

The inference of deep neural networks (DNNs) on resource-constrained embedded systems introduces non-trivial trade-offs among model accuracy, computational latency, and hardware limitations, particularly when real-time constraints must be satisfied. This paper presents ALADIN, an accuracy-latency-aware design-space inference analysis framework for mixed-precision quantized neural networks (QNNs) targeting scratchpad-based AI accelerators. ALADIN enables the evaluation and analysis of inference bottlenecks and design trade-offs across accuracy, latency, and resource consumption without requiring deployment on the target platform, thereby significantly reducing development time and cost. The framework introduces a progressive refinement process that transforms a canonical QONNX model into platform-aware representations by integrating both platform-independent implementation details and hardware-specific characteristics. ALADIN is validated using a cycle-accurate simulator of a RISC-V based platform specialized for AI workloads, demonstrating its effectiveness as a tool for quantitative inference analysis and hardware-software co-design. Experimental results highlight how architectural decisions and mixed-precision quantization strategies impact accuracy, latency, and resource usage, and show that these effects can be precisely evaluated and compared using ALADIN, while also revealing subtle optimization tensions.

Quantized neural networks are well known for reducing the latency, power consumption, and model size without significant harm to the performance. This makes them highly appropriate for systems with limited resources and low power capacity. Mixed-precision quantization offers better utilization of customized hardware that supports arithmetic operations at different bitwidths. Quantization methods either aim to minimize the compression loss given a desired reduction or optimize a dependent variable for a specified property of the model (such as FLOPs or model size); both make the performance inefficient when deployed on specific hardware, but more importantly, quantization methods assume that the loss manifold holds a global minimum for a quantized model that copes with the global minimum of the full precision counterpart. Challenging this assumption, we argue that the optimal minimum changes as the precision changes, and thus, it is better to look at quantization as a random process, placing the foundation for a different approach to quantize neural networks, which, during the training procedure, quantizes the model to a different precision, looks at the bit allocation as a Markov Decision Process, and then, finds an optimal bitwidth allocation for measuring specified behaviors on a specific device via direct signals from the particular hardware architecture. By doing so, we avoid the basic assumption that the loss behaves the same way for a quantized model. Automatic Mixed-Precision Quantization for Edge Devices (dubbed AMED) demonstrates its superiority over current state-of-the-art schemes in terms of the trade-off between neural network accuracy and hardware efficiency, backed by a comprehensive evaluation.

A suitable framework for the development of artificial neural networks is important because it decides the level of accuracy, which can be reached for a certain dataset and increases the certainty about the reached classification results. In this paper, we conduct a comparative study for the performance of four frameworks, Keras with TensorFlow, Pytorch, TensorFlow, and Cognitive Toolkit (CNTK), for the elaboration of neural networks. The number of neurons in the hidden layer of the neural networks is varied from 8 to 64 to understand its effect on the performance metrics of the frameworks. A test dataset is synthesized using an analytical model and real measured impedance spectra by an eddy current sensor coil on EUR 2 and TRY 1 coins. The dataset has been extended by using a novel method based on interpolation technique to create datasets with different difficulty levels to replicate the scenario with a good imitation of EUR 2 coins and to investigate the limit of the prediction accuracy. It was observed that the compared frameworks have high accuracy performance for a lower level of difficulty in the dataset. As the difficulty in the dataset is raised, there was a drop in the accuracy of CNTK and Keras with TensorFlow depending upon the number of neurons in the hidden layers. It was observed that CNTK has the overall worst accuracy performance with an increase in the difficulty level of the datasets. Therefore, the major comparison was confined to Pytorch and TensorFlow. It was observed for Pytorch and TensorFlow with 32 and 64 neurons in hidden layers that there is a minor drop in the accuracy with an increase in the difficulty level of the dataset and was above 90% until both the coins were 80% closer to each other in terms of electrical and magnetic properties. However, Pytorch with 32 neurons in the hidden layer has a reduction in model size by 70% and 16.3% and predicts the class, 73.6% and 15.6% faster in comparison to TensorFlow and Pytorch with 64 neurons.

Binary convolutional neural networks (BCNN) have shown good accuracy for small to medium neural network models. Their extreme quantization of weights and activations reduces off-chip data transfer and greatly reduces the computational complexity of convolutions. Further reduction in the complexity of a BCNN model for fast execution can be achieved with model size reduction at the cost of network accuracy. In this paper, a multi-model inference technique is proposed to reduce the execution time of the binarized inference process without accuracy reduction. The technique considers a cascade of neural network models with different computation/accuracy ratios. A parameterizable binarized neural network with different trade-offs between complexity and accuracy is used to obtain multiple network models. We also propose a hardware accelerator to run multi-model inference throughput in embedded systems. The multi-model inference accelerator is demonstrated on low-density Zynq-7010 and Zynq-7020 FPGA devices, classifying images from the CIFAR-10 dataset. The proposed accelerator improves the frame rate per number of LUTs by 7.2× those of previous solutions on a ZYNQ7020 FPGA with similar accuracy. This shows the effectiveness of the multi-model inference technique and the efficiency of the proposed hardware accelerator.

The ultra-wideband, frequency agile and high duty cycle features of Low Probability of Intercept (LPI) radar pose serious concerns for passive radar detection systems that must efficiently identify LPI radar modulation schemes with very low latency and high accuracy. For these tasks, conventional deep learning techniques have often been employed, though this comes at the cost of increased model complexity and high energy consumption when implemented in tactically relevant scenarios with significant channel deterioration and size, weight, and power (SWaP) constrained platforms. In this paper, we propose a neuromorphic approach based on energy-efficient and robust Spiking Neural Networks (SNNs). To demonstrate the effectiveness of our proposed approach, we perform a challenging LPI radar waveform recognition task with a Spiking Convolutional Neural Network (SCNN). Simulation results show that the SCNN achieves a 13-waveform recognition accuracy of 99.1% at OdB SNR with 1.51µJ of total energy per inference. It is found that our system outperforms non-spiking models with a reduction in model size and three orders of magnitude lower energy consumption. To the best of our knowledge, this work presents the first spike-based implementation of an LPI radar waveform recognition system. Moreover, this paper highlights the potential for extremely low SWaP neuromorphic hardware deployment to enable intelligent electronic support systems that demand high throughput efficiency and ultra-low power.

It is common knowledge that the optimal values of the calibrated parameters of a rainfall-runoff model for one model response may not be the optimal values for another model response. Thus, it is highly desirable to derive a Pareto front or trade-off curve on which each point represents a set of optimal values satisfying the desirable accuracy levels of each of the model responses. This paper presents a new genetic algorithm (GA) based calibration scheme, accelerated convergence GA (ACGA), which generates a limited number of points on the Pareto front. A neural network (NN) is then trained to compliment ACGA in the derivation of other desired points on the Pareto front by mimicking the relationship between the ACGA-generated calibration parameters and the model responses. The calibration scheme, ACGA, is linked with HydroWorks and tested on a catchment in Singapore. Results show that ACGA is more efficient and effective in deriving the Pareto front compared to other established GA-based optimization techniques such as vector evaluated GA, multiobjective GA, and nondominated sorting GA. Verification of the trained NN forecaster indicates that the trained network reproduces ACGA generated points on the Pareto front accurately. Thus, ACGA-NN is a useful and reliable tool to generate additional points on the Pareto front.

Convolutional neural network (CNN) accelerators have achieved great success from cloud to edge scenarios. However, given the trend towards even larger and deeper neural network models, it remains a challenging problem to efficiently process these CNNs especially on edge devices with limited energy budget. Accordingly, reducing the energy consumption is of paramount importance for sustainable CNN accelerators. In this paper, we propose AdaPrune, a novel pruning technique that reduces model size and computation to achieve performance improvement and energy savings for CNN accelerators. Unlike previous pruning techniques that sacrifice either computational regularity or accuracy, AdaPrune maintains both by customizing CNN pruning for the underlying accelerators to maximally leverage the sparsity benefits. AdaPrune consists of two techniques: input channel group pruning and output channel group pruning. By analyzing the weight fetching patterns of sparse CNN accelerators, AdaPrune adaptively switches between the two techniques to guarantee that the zeros are evenly distributed in each fetching group. In doing so, the pruned network structure preserves customized computational regularity for the underlying accelerators, thereby boosting the performance and energy efficiency. We evaluate AdaPrune on three sparse CNN accelerators with different spatial tiling strategies. The experimental results show that AdaPrune achieves up to 1.6× performance speedup, and 1.5× energy savings compared to unstructured pruning.

The Deep Learning (DL) approach has been gaining much popularity in recent years in the development of electroencephalogram (EEG) based Motor Imagery (MI) Brain-Computer Interface (BCI) systems, aiming to improve the performance of existing stroke rehabilitation strategies. A complex deep neural network structure has lots of neurons with thousands of parameters to optimize, and a great deal of data is often required to train the network and the training process can take an extremely long time. High training costs and high model complexity not only have negative impacts on the performance of the BCI system but also on its applicability to meet the real-time requirement to support the rehabilitation exercises of patients. To tackle the challenge, a contribution-based neuron selection method is proposed in this paper. A Convolutional Neural Network (CNN) based motor imagery classification framework is implemented, and a neuron pruning approach is developed and applied. The temporal and spatial features of EEG signals are captured by the CNN layers, and then the fast recursive algorithm (FRA) is applied to prune redundant parameters in the fully connected layers which reduces the computation cost of the CNN model without affecting its performance. The experimental results show that the proposed method can achieve up to 50% model size reduction and 67.09% computation savings.

Convolutional Neural Networks also known as ConvNets are now extensively used in various machine learning tasks for solving various problems in computer vision, biomedical, defence industry, entertainment etc. These neural networks for most of the applications are focused towards increasing the accuracy. However, besides maintaining the accuracy within a tolerable range, reduction in the network model size can have a lot of advantages from its mobility, easy deployment, remote upgradation and energy efficiency point of view. To attain these advantages, we propose a universal strategy to realize the convolution operation of a n x n filter kernel with fewer parameters, which also reduces the number of channels. We have proposed a compressed VGGNet model based on VGGNet neural network which resulted in 20x lesser parameters compared to its classical counterpart with an improved inference time by 3 times whilst maintaining similar accuracy. A complete hardware design of the compressed VGG architecture has also been implemented. A quantitative and qualitative analysis for various variants of VGGNet and other models reveal the reduction in the number of parameters in the range of 18-20x and the number of network operations contributing to the model complexity has shown a reduction of 2.5x with respect to its vanilla counterpart making it easier to deploy onto FPGAs and edge devices

A multilayer feedforward neural network with a modification to the standard backpropagation training of neural nets is considered. Information on the gradients of target outputs with respect to network inputs is used in this modification. This modified neural network is then applied for approximating design requirements of aerospace composite components, such as an aircraft engine guide vane and a satellite reflector assembly. This neural network functional approximator is then used with an optimization algorithm for determining the optimal parameters for component design requirements. The modified neural network provides for faster convergence of learning process and requires a smaller number of hidden layer nodes for convergence to similar training error levels. HE application of neural networks as universal approximators has been the focus of tremendous amount of activity in the past few years.16 Neural networks have been used for approximating and modeling both memoryless systems and dynamical systems with memory. This paper does not go into the various neural net architectures and training schemes, and the interested reader could refer to standard texts on this subject matter.1'2 This paper considers static neural networks (systems without memory) formed by interconnections without loops, called feedforward networks.2 In addition, this work considers a backpropagation learning scheme with adaptation of learning rate. There have been many additions to, or adaptations of, the basic backpropagation scheme, such as adaptive learning rate, learning with momentum, and the like,1'2 for improving convergence of the learning scheme, and these are not addressed in this paper. This paper explores a modification of the backpropagation learning scheme using information on the gradient of the target outputs with respect to the inputs of the network. The performance of the modified learning scheme is compared with that of the standard backpropagation learning scheme.2 A feedforward network with a single hidden layer trained with this modified learning scheme is applied in structural optimization problems. Figure 1 shows a feedforward net with a single hidden layer. The restriction to a single hidden layer is not a limitation of the modified learning scheme, and the new training scheme can be used on networks with multiple hidden layers also. The use of neural networks for structural response approximation has been reported by several authors.46 The primary reason for investigating such properties is to assess the applicability of trained neural networks for quick and accurate approximation of response functions during structural design optimization. Structural optimization, utilizing either gradient-based numerical search or genetic search algorithms, requires repeated function (response) evaluations during the search process. A single finite element analysis of large-scale structures requires a significant amount of computational time and resource and hence the need for robust approxima

Convolutional neural networks (CNNs) provide the best accuracy for disparity estimation. However, CNNs are computationally expensive, making them unfavorable for resource-limited devices with real-time constraints. Recent advances in neural architectures search (NAS) promise opportunities in automated optimization for disparity estimation. However, the main challenge of the NAS methods is the significant amount of computing time to explore a vast search space [e.g., <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.6\times 10^{29}$ </tex-math></inline-formula> ] and costly training candidates. To reduce the NAS computational demand, many proxy-based NAS methods have been proposed. Despite their success, most of them are designed for comparatively small-scale learning tasks. In this article, we propose a fast NAS method, called FastStereoNet, to enable resource-aware NAS within an intractably large search space. FastStereoNet automatically searches for hardware-friendly CNN architectures based on late acceptance hill climbing (LAHC), followed by simulated annealing (SA). FastStereoNet also employs a fine-tuning with a transferred weights mechanism to improve the convergence of the search process. The collection of these ideas provides competitive results in terms of search time and strikes a balance between accuracy and efficiency. Compared to the state of the art, FastStereoNet provides <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$5.25\times $ </tex-math></inline-formula> reduction in search time and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$44.4\times $ </tex-math></inline-formula> reduction in model size. These benefits are attained while yielding a comparable accuracy that enables seamless deployment of disparity estimation on resource-limited devices. Finally, FastStereoNet significantly improves the perception quality of disparity estimation deployed on field-programmable gate array and Intel Neural Compute Stick 2 accelerator in a significantly less onerous manner.