Weights Of Deep Neural Network Research Articles

Accelerated Gradient Approach For Deep Neural Network-Based Adaptive Control of Unknown Nonlinear Systems.

Recent connections in the adaptive control literature to continuous-time analogs of Nesterov's accelerated gradient method have led to the development of new real-time adaptation laws based on accelerated gradient methods. However, previous results assume that the system's uncertainties are linear-in-the-parameters (LIP). To compensate for non-LIP uncertainties, our preliminary results developed a neural network (NN)-based accelerated gradient adaptive controller to achieve trajectory tracking for nonlinear systems; however, the development and analysis only considered single-hidden-layer NNs. In this article, a generalized deep NN (DNN) architecture with an arbitrary number of hidden layers is considered, and a new DNN-based accelerated gradient adaptation scheme is developed to generate estimates of all the DNN weights in real-time. A nonsmooth Lyapunov-based analysis is used to guarantee the developed accelerated gradient-based DNN adaptation design achieves global asymptotic tracking error convergence for general nonlinear control affine systems subject to unknown (non-LIP) drift dynamics and exogenous disturbances. A comprehensive set of simulation studies are conducted on a two-state nonlinear system, a robotic manipulator, and a complex 20-D nonlinear system to demonstrate the improved performance of the developed method. Our simulation studies demonstrate enhanced tracking and function approximation performance from both DNN architectures and accelerated gradient adaptation.

Investigation in optimization of process parameters in turning of mild steel using response surface methodology and modified deep neural network

The turning operation is a traditional and conventional process for machining cylindrical materials to achieve a desired shape. During machining in the turning process, the accuracy and quality of the end product mainly depend on its process parameters. Therefore, this study attempts to achieve a quality product from a conventional lathe machine by optimizing its process parameters for En2-BS970 mild steel. The experiments were designed with the help of response surface methodology (RSM). The most influential turning parameters, such as feed rate, spindle speed, cutting fluid flow rate, and cutting angle, are investigated experimentally. In addition, the machining responses, namely material removal rate (MRR), surface roughness (SR), cutting force (CF), and cutting time (CT), have been optimized using the RSM numerical optimization method. Furthermore, a deep neural network (DNN) machine learning prediction model based on the whale optimization algorithm (WOA) is developed for this experiment in order to predict turning performances for En2-BS970 material. The predicted DNN results showed an accuracy of about 90% compared to experimental results, indicating that the implementation of WOA significantly optimized the DNN weights during training. The RSM optimized responses are obtained as 14.608 mm3/min for MRR, 0.7504 µm for SR, 442.94 N for CF, and 2.48 seconds for CT during the turning operation with input settings of 0.24 mm/rev, 466.535 rpm, 0.075 ml/s and 60.18° for feed rate, spindle speed, cutting fluid flow rate and cutting edge angle respectively.

Smart-DNN+: A Memory-efficient Neural Networks Compression Framework for the Model Inference

Deep Neural Networks (DNNs) have achieved remarkable success in various real-world applications. However, running a Deep Neural Network (DNN) typically requires hundreds of megabytes of memory footprints, making it challenging to deploy on resource-constrained platforms such as mobile devices and IoT. Although mainstream DNNs compression techniques such as pruning, distillation, and quantization can reduce the memory overhead of model parameters during DNN inference, they suffer from three limitations: (i) low model compression ratio for the lightweight DNN structures with little redundancy, (ii) potential degradation in model inference accuracy, and (iii) inadequate memory compression ratio is attributable to ignoring the layering property of DNN inference. To address these issues, we propose a lightweight memory-efficient DNN inference framework called Smart-DNN+, which significantly reduces the memory costs of DNN inference without degrading the model quality. Specifically, ① Smart-DNN+ applies a layerwise binary-quantizer with a remapping mechanism to greatly reduce the model size by quantizing the typical floating-point DNN weights of 32-bit to the 1-bit signs layer by layer. To maintain model quality, ② Smart-DNN+ employs a bucket-encoder to keep the compressed quantization error by encoding the multiple similar floating-point residuals into the same integer bucket IDs. When running the compressed DNN in the user’s device, ③ Smart-DNN+ utilizes a partially decompressing strategy to greatly reduce the required memory overhead by first loading the compressed DNNs in memory and then dynamically decompressing the required materials for model inference layer by layer. Experimental results on popular DNNs and datasets demonstrate that Smart-DNN+ achieves lower 0.17%–0.92% memory costs at lower runtime overheads compared with the states of the art without degrading the inference accuracy. Moreover, Smart-DNN+ potentially reduces the inference runtime up to 2.04× that of conventional DNN inference workflow.

SmartDeal: Remodeling Deep Network Weights for Efficient Inference and Training.

The record-breaking performance of deep neural networks (DNNs) comes with heavy parameter budgets, which leads to external dynamic random access memory (DRAM) for storage. The prohibitive energy of DRAM accesses makes it nontrivial for DNN deployment on resource-constrained devices, calling for minimizing the movements of weights and data in order to improve the energy efficiency. Driven by this critical bottleneck, we present SmartDeal, a hardware-friendly algorithm framework to trade higher-cost memory storage/access for lower-cost computation, in order to aggressively boost the storage and energy efficiency, for both DNN inference and training. The core technique of SmartDeal is a novel DNN weight matrix decomposition framework with respective structural constraints on each matrix factor, carefully crafted to unleash the hardware-aware efficiency potential. Specifically, we decompose each weight tensor as the product of a small basis matrix and a large structurally sparse coefficient matrix whose nonzero elements are readily quantized to the power-of-2. The resulting sparse and readily quantized DNNs enjoy greatly reduced energy consumption in data movement as well as weight storage, while incurring minimal overhead to recover the original weights thanks to the required sparse bit-operations and cost-favorable computations. Beyond inference, we take another leap to embrace energy-efficient training, by introducing several customized techniques to address the unique roadblocks arising in training while preserving the SmartDeal structures. We also design a dedicated hardware accelerator to fully utilize the new weight structure to improve the real energy efficiency and latency performance. We conduct experiments on both vision and language tasks, with nine models, four datasets, and three settings (inference-only, adaptation, and fine-tuning). Our extensive results show that 1) being applied to inference, SmartDeal achieves up to 2.44x improvement in energy efficiency as evaluated using real hardware implementations and 2) being applied to training, SmartDeal can lead to 10.56x and 4.48x reduction in the storage and the training energy cost, respectively, with usually negligible accuracy loss, compared to state-of-the-art training baselines. Our source codes are available at: https://github.com/VITA-Group/SmartDeal.

Booth Encoding-Based Energy Efficient Multipliers for Deep Learning Systems

Artificial intelligence on edge is a growing research field. In this paper, we propose a novel re-encoding scheme for reducing the size of the weights of deep neural networks (DNNs). The proposed re-encoding scheme exploits the Booth encoding scheme and the power-of-two (PO2) quantization to allow for very low energy computations during the inference of the neural networks with minimal loss in classification accuracy. We demonstrate the advantages of the proposed re-encoding scheme by computing a convolutional neural network (CNN) and a linear neural network on the proposed Extended Exact Multiplier and the proposed PO2 Multiplier. Our proposed PO2 quantization and re-encoding method reduce the model size for the CNN by 30.77% and the model size of the linear neural network by 49.86%. Furthermore, our multipliers reduce the inference energy for CNN by 50.6% and for the linear neural network by 90.1%. The PO2 Multiplier is proposed for the sensor-end computation of the linear neural network with a 77.32% reduction in the area relative to an exact Booth multiplier and it reduces the inference energy consumption of the linear neural network by 93.2% when compared to the unmodified exact multiplier. Our proposed scheme can be used to improve the energy consumption during inference for most Booth multipliers with minor modifications to the re-encoding signal arrangements. We also demonstrate that the proposed re-encoding scheme paired with the proposed multipliers outperforms all the existing designs in terms of resource utilization with a minimal impact on the inference accuracy of the neural networks.

Robust Deep Learning for Uplink Channel Estimation in Cellular Network Under Inter-Cell Interference

Deep learning (DL)-based channel estimation has achieved remarkable success. However, most existing works focus on the white Gaussian noise which are inapplicable for cell-edge users under inter-cell interference (ICI). In this paper, we address this issue by proposing a novel DL-based channel estimation solution with a cascaded model-based and model-free deep neural network (DNN) structure. Specifically, the model-based module is designed by the variational Bayesian inference (VBI) technique to suppress the time-varying ICI, and the model-free module is designed by the Denoising Sparse Autoencoder (DSAE) structure to further refine the channel estimation. The proposed DNN is firstly pre-trained by offline supervised training, and various channel statistics are encapsulated in the DNN weights with the assist of a hyper-prior net modelling different sparse priors for different training samples. Then, an online Bayesian learning algorithm is proposed to train the model-based VBI module based on real-time pilot samples to track the online channel statistics. Simulation results verify that the proposed solution outperforms various state-of-the-art baseline schemes in a large SINR range with comparable performance to the estimator with genie-aided channel statistics.

Deep Neural Network-Based Approximate Optimal Tracking for Unknown Nonlinear Systems

The infinite horizon optimal tracking problem is solved for a deterministic, control-affine, unknown nonlinear dynamical system. A deep neural network (DNN) is updated in real-time to approximate the unknown nonlinear system dynamics. The developed framework uses a multi-timescale concurrent learning-based weight update policy, with which the output-layer DNN weights are updated in real-time, but the internal DNN features are updated discretely and at a slower timescale (i.e., with batch-like updates). The design of the output-layer weight update policy is motivated by a Lyapunov-based analysis, and the inner features are updated according to existing DNN optimization algorithms. Simulation results demonstrate the efficacy of the developed technique and compare its performance to existing techniques.

Random and Adversarial Bit Error Robustness: Energy-Efficient and Secure DNN Accelerators.

Deep neural network (DNN) accelerators received considerable attention in recent years due to the potential to save energy compared to mainstream hardware. Low-voltage operation of DNN accelerators allows to further reduce energy consumption, however, causes bit-level failures in the memory storing the quantized weights. Furthermore, DNN accelerators are vulnerable to adversarial attacks on voltage controllers or individual bits. In this paper, we show that a combination of robust fixed-point quantization, weight clipping, as well as random bit error training (RandBET) or adversarial bit error training (AdvBET) improves robustness against random or adversarial bit errors in quantized DNN weights significantly. This leads not only to high energy savings for low-voltage operation as well as low-precision quantization, but also improves security of DNN accelerators. In contrast to related work, our approach generalizes across operating voltages and accelerators and does not require hardware changes. Moreover, we present a novel adversarial bit error attack and are able to obtain robustness against both targeted and untargeted bit-level attacks. Without losing more than 0.8%/2% in test accuracy, we can reduce energy consumption on CIFAR10by 20%/30% for 8/4-bit quantization. Allowing up to 320 adversarial bit errors, we reduce test error from above 90% (chance level) to 26.22%.

Knowledge Sharing in AI Services: A Market-Based Approach

Today’s deep neural networks (DNNs) are very accurate when trained on a large amount of data. However, suitable input might not be available or may require extensive data collection. Data sharing is one option to address these issues, but it is generally impractical because of privacy concerns or due to the problematic process of finding a sharing agreement. Instead, this work considers knowledge sharing by first exchanging the weights of pretrained DNNs and then applying transfer learning (TL). Specifically, it addresses the economics of knowledge sharing in AI services by taking a market-based approach. In detail, a model based on Fisher’s market is devised for optimal knowledge sharing, defined as the gain in inference accuracy from exchanging DNN weights. The proposed approach is shown to reach a market equilibrium and to satisfy important economic properties, including Pareto optimality. A technique for weight fusion is also introduced to merge acquired knowledge with the existing one. Finally, an extensive evaluation is conducted in a distributed intelligence scenario. The obtained results show that the proposed solution is efficient and that weight fusion with TL significantly increases inference accuracy compared to the original DNN, without the overhead of federated learning.

Robust Speaker Verification Using Deep Weight Space Ensemble

Domain shift is one of the most challenging problems in speaker verification. Although numerous methods have been proposed to address domain shift, most approaches optimize the performance of one domain at the sacrifice of the other. As a result, to obtain the best performance, each domain requires a dedicated model. However, deploying multiple models is resource-demanding and impractical, particularly when the deployment domains are not known in advance. Recent studies in deep neural networks (DNNs) suggest that near the low error surface of the DNN's weight space, there exists a linear path connecting a base model and a fine-tuned model. This finding inspires us to combine the strength of the fine-tuned models and the base models to solve challenging SV problems. Specifically, we aim to develop models that can handle 1) mixed text-dependent (TD) and text-independent (TI) speaker verification where the speech content can be either unconstrained or constrained, 2) cross-channel speaker verification where the recording can be 16 kHz high-fidelity microphone speech or 8 kHz telephone speech, and 3) bi-lingual speaker verification where the enrollment and test speech can be one of the two languages. With weight space ensemble, we show that we can substantially improve the tasks mentioned above, with a 39.6% improvement in mixing TD and TI SV, a 17.4% improvement in bi-lingual SV, and an 18.4% improvement in cross-channel SV. Moreover, we show that the weight space ensemble can also enhance the performance in the target domain, thanks to the regularization effect of the interpolation.

A Memory-Efficient Learning Framework for Symbol Level Precoding With Quantized NN Weights

This paper proposes a memory-efficient deep neural network (DNN) framework-based symbol level precoding (SLP). We focus on a DNN with realistic finite precision weights and adopt an unsupervised deep learning (DL) based SLP model (SLP-DNet). We apply a stochastic quantization (SQ) technique to obtain its corresponding quantized version called SLP-SQDNet. The proposed scheme offers a scalable performance vs memory trade-off, by quantizing a scalable percentage of the DNN weights, and we explore binary and ternary quantizations. Our results show that while SLP-DNet provides near-optimal performance, its quantized versions through SQ yield ~3.46× and ~2.64× model compression for binary-based and ternary-based SLP-SQDNets, respectively. We also find that our proposals offer ~20× and ~10× computational complexity reductions compared to SLP optimization-based and SLP-DNet, respectively.

AccHashtag : Accelerated Hashing for Detecting Fault-Injection Attacks on Embedded Neural Networks

We propose AccHashtag , the first framework for high-accuracy detection of fault-injection attacks on Deep Neural Networks (DNNs) with provable bounds on detection performance. Recent literature in fault-injection attacks shows the severe DNN accuracy degradation caused by bit flips. In this scenario, the attacker changes a few DNN weight bits during execution by injecting faults to the dynamic random-access memory (DRAM). To detect bit flips, AccHashtag extracts a unique signature from the benign DNN prior to deployment. The signature is used to validate the model’s integrity and verify the inference output on the fly. We propose a novel sensitivity analysis that identifies the most vulnerable DNN layers to the fault-injection attack. The DNN signature is constructed by encoding the weights in vulnerable layers using a low-collision hash function. During DNN inference, new hashes are extracted from the target layers and compared against the ground-truth signatures. AccHashtag incorporates a lightweight methodology that allows for real-time fault detection on embedded platforms. We devise a specialized compute core for AccHashtag on field-programmable gate arrays (FPGAs) to facilitate online hash generation in parallel to DNN execution. Extensive evaluations with the state-of-the-art bit-flip attack on various DNNs demonstrate the competitive advantage of AccHashtag in terms of both attack detection and execution overhead.

T-BFA: Targeted Bit-Flip Adversarial Weight Attack.

Traditional Deep Neural Network (DNN) security is mostly related to the well-known adversarial input example attack.Recently, another dimension of adversarial attack, namely, attack on DNN weight parameters, has been shown to be very powerful. Asa representative one, the Bit-Flip based adversarial weight Attack (BFA) injects an extremely small amount of faults into weight parameters to hijack the executing DNN function. Prior works of BFA focus on un-targeted attacks that can hack all inputs into a random output class by flipping a very small number of weight bits stored in computer memory. This paper proposes the first work oftargetedBFA based (T-BFA) adversarial weight attack on DNNs, which can intentionally mislead selected inputs to a target output class. The objective is achieved by identifying the weight bits that are highly associated with classification of a targeted output through a class-dependent weight bit searching algorithm. Our proposed T-BFA performance is successfully demonstrated on multiple DNN architectures for image classification tasks. For example, by merely flipping 27 out of 88 million weight bits of ResNet-18, our T-BFA can misclassify all the images from Hen class into Goose class (i.e., 100% attack success rate) in ImageNet dataset, while maintaining 59.35% validation accuracy.

GRIM: A General, Real-Time Deep Learning Inference Framework for Mobile Devices Based on Fine-Grained Structured Weight Sparsity.

It is appealing but challenging to achieve real-time deep neural network (DNN) inference on mobile devices, because even the powerful modern mobile devices are considered as "resource-constrained" when executing large-scale DNNs. It necessitates the sparse model inference via weight pruning, i.e., DNN weight sparsity, and it is desirable to design a new DNN weight sparsity scheme that can facilitate real-time inference on mobile devices while preserving a high sparse model accuracy. This paper designs a novel mobile inference acceleration framework GRIM that is General to both convolutional neural networks (CNNs) and recurrent neural networks (RNNs) and that achieves Real-time execution and high accuracy, leveraging fine-grained structured sparse model Inference and compiler optimizations for Mobiles. We start by proposing a new fine-grained structured sparsity scheme through the Block-based Column-Row (BCR) pruning. Based on this new fine-grained structured sparsity, our GRIM framework consists of two parts: (a) the compiler optimization and code generation for real-time mobile inference; and (b) the BCR pruning optimizations for determining pruning hyperparameters and performing weight pruning. We compare GRIM with Alibaba MNN, TVM, TensorFlow-Lite, a sparse implementation based on CSR, PatDNN, and ESE (a representative FPGA inference acceleration framework for RNNs), and achieve up to 14.08× speedup.

Multistrategy Improved Sparrow Search Algorithm Optimized Deep Neural Network for Esophageal Cancer.

Deep neural network is a complex pattern recognition network system. It is widely favored by scholars for its strong nonlinear fitting ability. However, training deep neural network models on small datasets typically realizes worse performance than shallow neural network. In this study, a strategy to improve the sparrow search algorithm based on the iterative map, iterative perturbation, and Gaussian mutation is developed. This optimized strategy improved the sparrow search algorithm validated by fourteen benchmark functions, and the algorithm has the best search accuracy and the fastest convergence speed. An algorithm based on the iterative map, iterative perturbation, and Gaussian mutation improved sparrow search algorithm is designed to optimize deep neural networks. The modified sparrow algorithm is exploited to search for the optimal connection weights of deep neural network. This algorithm is implemented for the esophageal cancer dataset along with the other six algorithms. The proposed model is able to achieve 0.92 under all the eight scoring criteria, which is better than the performance of the other six algorithms. Therefore, an optimized deep neural network based on an improved sparrow search algorithm with iterative map, iterative perturbation, and Gaussian mutation is an effective approach to predict the survival rate of esophageal cancer.

Perturbation of deep autoencoder weights for model compression and classification of tabular data

Fully connected deep neural networks (DNN) often include redundant weights leading to overfitting and high memory requirements. Additionally, in tabular data classification, DNNs are challenged by the often superior performance of traditional machine learning models. This paper proposes periodic perturbations (prune and regrow) of DNN weights, especially at the self-supervised pre-training stage of deep autoencoders. The proposed weight perturbation strategy outperforms dropout learning or weight regularization (L1 or L2) for four out of six tabular data sets in downstream classification tasks. Unlike dropout learning, the proposed weight perturbation routine additionally achieves 15% to 40% sparsity across six tabular data sets, resulting in compressed pretrained models. The proposed pretrained model compression improves the accuracy of downstream classification, unlike traditional weight pruning methods that trade off performance for model compression. Our experiments reveal that a pretrained deep autoencoder with weight perturbation can outperform traditional machine learning in tabular data classification, whereas baseline fully-connected DNNs yield the worst classification accuracy. However, traditional machine learning models are superior to any deep model when a tabular data set contains uncorrelated variables. Therefore, the performance of deep models with tabular data is contingent on the types and statistics of constituent variables.

Non-Structured DNN Weight Pruning-Is It Beneficial in Any Platform?

Large deep neural network (DNN) models pose the key challenge to energy efficiency due to the significantly higher energy consumption of off-chip DRAM accesses than arithmetic or SRAM operations. It motivates the intensive research on model compression with two main approaches. Weight pruning leverages the redundancy in the number of weights and can be performed in a non-structured, which has higher flexibility and pruning rate but incurs index accesses due to irregular weights, or structured manner, which preserves the full matrix structure with a lower pruning rate. Weight quantization leverages the redundancy in the number of bits in weights. Compared to pruning, quantization is much more hardware-friendly and has become a "must-do" step for FPGA and ASIC implementations. Thus, any evaluation of the effectiveness of pruning should be on top of quantization. The key open question is, with quantization, what kind of pruning (non-structured versus structured) is most beneficial? This question is fundamental because the answer will determine the design aspects that we should really focus on to avoid the diminishing return of certain optimizations. This article provides a definitive answer to the question for the first time. First, we build ADMM-NN-S by extending and enhancing ADMM-NN, a recently proposed joint weight pruning and quantization framework, with the algorithmic supports for structured pruning, dynamic ADMM regulation, and masked mapping and retraining. Second, we develop a methodology for fair and fundamental comparison of non-structured and structured pruning in terms of both storage and computation efficiency. Our results show that ADMM-NN-S consistently outperforms the prior art: 1) it achieves 348× , 36× , and 8× overall weight pruning on LeNet-5, AlexNet, and ResNet-50, respectively, with (almost) zero accuracy loss and 2) we demonstrate the first fully binarized (for all layers) DNNs can be lossless in accuracy in many cases. These results provide a strong baseline and credibility of our study. Based on the proposed comparison framework, with the same accuracy and quantization, the results show that non-structured pruning is not competitive in terms of both storage and computation efficiency. Thus, we conclude that structured pruning has a greater potential compared to non-structured pruning. We encourage the community to focus on studying the DNN inference acceleration with structured sparsity.

Deep reinforcement learning-based decision support system for transportation infrastructure management under hurricane events

The high wind, large surge/wave and heavy rainfall during a hurricane may significantly impact the performance of transportation infrastructure and associated traffic safety. Accordingly, stakeholders need to make a sequence of decisions to close (or restrict) the traffic of vulnerable components in the transportation network (e.g., aerodynamics-sensitive long-span bridges, hydrodynamics-sensitive coastal bridges and inundation-sensitive road segments) for the balance of safety and mobility. Due to large uncertainty involved in hurricane weather and traffic conditions, it is essentially a stochastic sequential decision problem and can be effectively formulated as a Markov decision process. This study proposes a deep reinforcement learning (RL)-based decision support system for stakeholders to optimally manage these critical components for the purpose of minimizing the network-level losses induced by hurricanes. Specifically, the RL policy (i.e., mapping from high-dimensional continuous traffic/weather information to management decisions) is represented by a deep neural network (DNN) and the optimal decision (corresponding to a set of DNN weights) is obtained using the deep Q learning algorithm. A case study based on a hypothetical traffic network is utilized to demonstrate the good performance of the developed deep RL-based decision support system in the application of transportation infrastructure management under hurricane conditions.

A Heterogeneous In-Memory Computing Cluster for Flexible End-to-End Inference of Real-World Deep Neural Networks

Deployment of modern TinyML tasks on small battery-constrained IoT devices requires high computational energy efficiency. Analog In-Memory Computing (IMC) using non-volatile memory (NVM) promises major efficiency improvements in deep neural network (DNN) inference and serves as on-chip memory storage for DNN weights. However, IMC’s functional flexibility limitations and their impact on performance, energy, and area efficiency are not yet fully understood at the system level. To target practical end-to-end IoT applications, IMC arrays must be enclosed in heterogeneous programmable systems, introducing new system-level challenges which we aim at addressing in this work. We present a heterogeneous tightly-coupled clustered architecture integrating 8 RISC-V cores, an in-memory computing accelerator (IMA), and digital accelerators. We benchmark the system on a highly heterogeneous workload such as the Bottleneck layer from a MobileNetV2, showing <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$11.5\times $ </tex-math></inline-formula> performance and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$9.5\times $ </tex-math></inline-formula> energy efficiency improvements, compared to highly optimized parallel execution on the cores. Furthermore, we explore the requirements for end-to-end inference of a full mobile-grade DNN (MobileNetV2) in terms of IMC array resources, by scaling up our heterogeneous architecture to a multi-array accelerator. Our results show that our solution, on the end-to-end inference of the MobileNetV2, is one order of magnitude better in terms of execution latency than existing programmable architectures and two orders of magnitude better than state-of-the-art heterogeneous solutions integrating in-memory computing analog cores.

Achieving High In Situ Training Accuracy and Energy Efficiency with Analog Non-Volatile Synaptic Devices

On-device embedded artificial intelligence prefers the adaptive learning capability when deployed in the field, and thus in situ training is required. The compute-in-memory approach, which exploits the analog computation within the memory array, is a promising solution for deep neural network (DNN) on-chip acceleration. Emerging non-volatile memories are of great interest, serving as analog synapses due to their multilevel programmability. However, the asymmetry and nonlinearity in the conductance tuning remain grand challenges for achieving high in situ training accuracy. In addition, analog-to-digital converters at the edge of the memory array introduce quantization errors. In this work, we present an algorithm-hardware co-optimization to overcome these challenges. We incorporate the device/circuit non-ideal effects into the DNN propagation and weight update steps. By introducing the adaptive “momentum” in the weight update rule, in situ training accuracy on CIFAR-10 could approach its software baseline even under severe asymmetry/nonlinearity and analog-to-digital converter quantization error. The hardware performance of the on-chip training architecture and the overhead for adding “momentum” are also evaluated. By optimizing the backpropagation dataflow, 23.59 TOPS/W training energy efficiency (12× improvement compared to naïve dataflow) is achieved. The circuits that handle “momentum” introduce only 4.2% energy overhead. Our results show great potential and more relaxed requirements that enable emerging non-volatile memories for DNN acceleration on the embedded artificial intelligence platforms.