Drop In Accuracy Research Articles

FxP-QNet: A Post-Training Quantizer for the Design of Mixed Low-Precision DNNs With Dynamic Fixed-Point Representation

Deep neural networks (DNNs) have demonstrated their effectiveness in a wide range of computer vision tasks, with the state-of-the-art results obtained through complex and deep structures that require intensive computation and memory. Now-a-days, efficient model inference is crucial for consumer applications on resource-constrained platforms. As a result, there is much interest in the research and development of dedicated deep learning (DL) hardware to improve the throughput and energy efficiency of DNNs. Low-precision representation of DNN data-structures through quantization would bring great benefits to specialized DL hardware. However, the rigorous quantization leads to a severe accuracy drop. As such, quantization opens a large hyper-parameter space at bit-precision levels, the exploration of which is a major challenge. In this paper, we propose a novel framework referred to as the Fixed-Point Quantizer of deep neural Networks (FxP-QNet) that flexibly designs a mixed low-precision DNN for integer-arithmetic-only deployment. Specifically, the FxP-QNet gradually adapts the quantization level for each data-structure of each layer based on the trade-off between the network accuracy and the low-precision requirements. Additionally, it employs post-training self-distillation and network prediction error statistics to optimize the quantization of floating-point values into fixed-point numbers. Examining FxP-QNet on state-of-the-art architectures and the benchmark ImageNet dataset, we empirically demonstrate the effectiveness of FxP-QNet in achieving the accuracy-compression trade-off without the need for training. The results show that FxP-QNet-quantized AlexNet, VGG-16, and ResNet-18 reduce the overall memory requirements of their full-precision counterparts by 7.16x, 10.36x, and 6.44x with less than 0.95%, 0.95%, and 1.99% accuracy drop, respectively.

Divergence-Based Transferability Analysis for Self-Adaptive Smart Grid Intrusion Detection With Transfer Learning

Machine learning is a popular approach to security monitoring and intrusion detection in cyber-physical systems (CPS) like the smart grid. However, these highly dynamic CPS operating in open environments can result in significant data distribution divergence, which may require the adaptation of a learned model. While transfer learning has been an effective approach to retain the performance against the divergence, there is still limited work on a more fundamental question that can be called <i>transferability</i>: when should one apply transfer learning? To address this challenge, this paper proposes a divergence-based transferability analysis to decide whether to apply transfer learning and autonomically adapt learning-based intrusion detectors. This work first identifies three metrics used to measure the divergence between data distributions, and then explores the relation between detector&#x2019;s accuracy drop and divergence in extensive temporal, spatial, and spatiotemporal experiments. Two regression models are trained to approximate the divergence-accuracy relation and then used to predict an accuracy drop which determines whether to apply transfer learning. Finally, a state-of-the-art domain adversarial neural network (DANN) classifier is adopted as the transfer learning model. Datasets from real normal operation profiles and simulated attacks are used to validate the effectiveness of the proposed transferability analysis against variations in attack timing, locations, and both. In all three scenarios, the proposed analysis demonstrated high accuracy in predicting accuracy drop from the divergence, with an RMSE lower than 4.20&#x0025;, and the DANN can be timely triggered to achieve an accuracy improvement over 5.00&#x0025;.

A novel approach for context-aware sensor optimization in a smart home

Ambient Intelligence is an emerging technology that allows for reasoning of activities that are carried out in a smart environment. Smart sensors play a vital role in activity recognition tasks. Though a great number of sensors provide more pinpointed information about where and when the interaction in the environment occurred, addition of sensors also results in higher power consumption and greater cost constraints. The number of training samples used for activity recognition tasks also increases and becomes computationally expensive. Thus, to overcome these constraints sensor optimization in smart environment is carried out. Mapping of sensors to a user's activity becomes tedious when more than one activity is carried out at the same timestamp. Unambiguous sensor-activity relationship is derived with the help of spatial and temporal context represented by ontology model that captures real time sensor data. Since there exits spatial noise during activity annotation process, it is eliminated with help of pointwise mutual information score. Spatial and temporal data of an activity helps sensor optimization in multi resident living with concurrent activity occurrence. Inactive sensors are removed using mutual information score and redundant sensors are contextually paired based on where an activity happens. By doing so, only unique set of sensors are paired every time even when concurrent activity occurs. The quality of features derived is tested by testing the accuracy of the model by activity recognition task using Random Forest algorithm. Nearly 24% of sensors were identified as inactive in the smart environment. The accuracy of Human Activity Recognition (HAR) task remained constant even after removing the inactive sensors. Addition of spatial context helped to improve HAR with increased accuracy. Contextual pairing of redundant sensors resulted in minimal drop in the model's accuracy.

Analysis on Effective UAS Survey Conditions for Classification of Coastal Sediments

This study aims to introduce effective unmanned aerial system (UAS) survey conditions for coastal sediment classification, including muddy sand, sand, gravel, and shells in a tidal flat area. UAS images with resolutions ranging from 2 to 60 mm are used as an implication of survey altitudes. The UAS images are used for sediment classification using random forest (RF) and support vector machine (SVM) methods. The results showed that RF is more effective in sediment classification while the general accuracy pattern was similar. The accuracy decreased with lower spatial resolutions. Notably, there is a significant drop of accuracy with a resolution coarser than 40 mm. Considering the training data selection, classification accuracy, and survey efficiency, it is suggested that 40 mm UAS images would provide optimal condition with acceptable accuracy for coastal sediment classification using RF model. To gain higher accuracy, a lower flight altitude is required, which will elongate the survey time significantly. Given the fact that this study is the first approach to test various UAS survey conditions for coastal sediment classifications in a field condition; the methodology and findings of this study can serve as a guideline framework for future coastal UAS sediment mapping.

Block Switching: A Stochastic Approach for Deep Learning Security

Recent study of adversarial attacks has revealed the vulnerability of modern deep learning models. That is, subtly crafted perturbations of the input can make a trained network with high accuracy and produce arbitrary incorrect predictions, while maintaining imperceptible to human vision system. In this paper, we introduce Block Switching (BS), a defense strategy against adversarial attacks based on stochasticity. BS replaces a block of model layers with multiple parallel channels, and the active channel is randomly assigned in the run time, hence unpredictable to the adversary. We show empirically that BS leads to a more dispersed input gradient distribution and superior defense effectiveness compared with other stochastic defenses such as stochastic activation pruning. Compared to other defenses, BS is also characterized by the following features: (i) BS causes less test accuracy drop; (ii) BS is attack-independent; and (iii) BS is compatible with other defenses and can be used jointly with others.

HRel: Filter pruning based on High Relevance between activation maps and class labels

This paper proposes an Information Bottleneck theory based filter pruning method that uses a statistical measure called Mutual Information (MI). The MI between filters and class labels, also called Relevance, is computed using the filter’s activation maps and the annotations. The filters having High Relevance (HRel) are considered to be more important. Consequently, the least important filters, which have lower Mutual Information with the class labels, are pruned. Unlike the existing MI based pruning methods, the proposed method determines the significance of the filters purely based on their corresponding activation map’s relationship with the class labels. Architectures such as LeNet-5, VGG-16, ResNet-56, ResNet-110 and ResNet-50 are utilized to demonstrate the efficacy of the proposed pruning method over MNIST, CIFAR-10 and ImageNet datasets. The proposed method shows the state-of-the-art pruning results for LeNet-5, VGG-16, ResNet-56, ResNet-110 and ResNet-50 architectures. In the experiments, we prune 97.98%, 84.85%, 76.89%, 76.95%, and 63.99% of Floating Point Operation (FLOP)s from LeNet-5, VGG-16, ResNet-56, ResNet-110, and ResNet-50 respectively. The proposed HRel pruning method outperforms recent state-of-the-art filter pruning methods. Even after pruning the filters from convolutional layers of LeNet-5 drastically (i.e., from 20, 50 to 2, 3, respectively), only a small accuracy drop of 0.52% is observed. Notably, for VGG-16, 94.98% parameters are reduced, only with a drop of 0.36% in top-1 accuracy. ResNet-50 has shown a 1.17% drop in the top-5 accuracy after pruning 66.42% of the FLOPs. In addition to pruning, the Information Plane dynamics of Information Bottleneck theory is analyzed for various Convolutional Neural Network architectures with the effect of pruning. The code is available at https://github.com/sarvanichinthapalli/HRel.

Artificial intelligence algorithms to predict Italian real estate market prices

PurposeThe assessment of the Real Estate (RE) prices depends on multiple factors that traditional evaluation methods often struggle to fully understand. Housing prices, in particular, are the foundations for a better knowledge of the Built Environment and its characteristics. Recently, Machine Learning (ML) techniques, which are a subset of Artificial Intelligence, are gaining momentum in solving complex, non-linear problems like house price forecasting. Hence, this study deployed three popular ML techniques to predict dwelling prices in two cities in Italy.Design/methodology/approachAn extensive dataset about house prices is collected through API protocol in two cities in North Italy, namely Brescia and Varese. This data is used to train and test three most popular ML models, i.e. ElasticNet, XGBoost and Artificial Neural Network, in order to predict house prices with six different features.FindingsThe models' performance was evaluated using the Mean Absolute Error (MAE) score. The results showed that the artificial neural network performed better than the others in predicting house prices, with a MAE 5% lower than the second-best model (which was the XGBoost).Research limitations/implicationsAll the models had an accuracy drop in forecasting the most expensive cases, probably due to a lack of data.Practical implicationsThe accessibility and easiness of the proposed model will allow future users to predict house prices with different datasets. Alternatively, further research may implement a different model using neural networks, knowing that they work better for this kind of task.Originality/valueTo date, this is the first comparison of the three most popular ML models that are usually employed when predicting house prices.

TempDiff: Feature Map-Level CNN Sparsity Enhancement at Near-Zero Memory Overhead via Temporal Difference

The demand for distributed vision systems at an unceasingly larger scale requires the availability of sensor nodes able to execute high-level visual tasks (e.g., object detection for visual monitoring). Such tasks routinely entail the high computational cost of Convolutional Neural Networks (CNN), which conflicts with the tight power budget and memory capacity of sensor nodes at the edge. This paper introduces an approach to reduce the computational cost of object detection in CNN accelerators available in edge devices. The proposed approach induces additional feature map-level sparsity (i.e., computation skipping) at inference time by exploiting temporal correlation among frames. The proposed approach has an uncommonly favorable computation-memory tradeoff, as significant computation reduction is achieved at the cost of very small additional memory for the storage of intermediate features. As further benefits, no architectural changes or retraining are required, allowing immediate deployment in existing vision frameworks and suppressing the need for storing multiple models. Results show that the proposed TempDiff method achieves up to 37% computation reduction with 1.1% accuracy drop based on the SSD(VGG16) object detection network, under both VIRAT and ImageNet-VID datasets. Similarly, 18.3% (35.8%) computation reduction at 3.3% (3.2%) memory overhead, and 3.8% (6.8%) accuracy drop in YOLOv1 (VGG16) (SSD (VGG16)) is achieved under the CAMEL dataset. Furthermore up to 58% computation reduction with 2% accuracy drop and 3.7% memory overhead were achieved for YOLOv3-Tiny network under the ImageNet-VID dataset.

Self-Adaptive Approximate Mobile Deep Learning

Edge intelligence is currently facing several important challenges hindering its performance, with the major drawback being meeting the high resource requirements of deep learning by the resource-constrained edge computing devices. The most recent adaptive neural network compression techniques demonstrated, in theory, the potential to facilitate the flexible deployment of deep learning models in real-world applications. However, their actual suitability and performance in ubiquitous or edge computing applications has not, to this date, been evaluated. In this context, our work aims to bridge the gap between the theoretical resource savings promised by such approaches and the requirements of a real-world mobile application by introducing algorithms that dynamically guide the compression rate of a neural network according to the continuously changing context in which the mobile computation is taking place. Through an in-depth trace-based investigation, we confirm the feasibility of our adaptation algorithms in offering a scalable trade-off between the inference accuracy and resource usage. We then implement our approach on real-world edge devices and, through a human activity recognition application, confirm that it offers efficient neural network compression adaptation in highly dynamic environments. The results of our experiment with 21 participants show that, compared to using static network compression, our approach uses 2.18× less energy with only a 1.5% drop in the average accuracy of the classification.

Deep Convolutional Neural Networks Detect Tumor Genotype from Pathological Tissue Images in Gastrointestinal Stromal Tumors.

Simple SummaryIn this study, we established four convolutional neural network (DCNN) models (AlexNet, ResNet101, DenseNet201, and InceptionV3) to predict drug-sensitive mutations from images of tissues with gastrointestinal stromal tumors. The treatment of these tumors depends on the mutational subtype of the KIT/PDGFRA genes. Previous studies rarely focused on mesenchymal tumors and mutational subtypes. More than 5000 images of 365 GISTs from three independent laboratories were used to generate the model. DenseNet201 achieved an accuracy of 87% while the accuracies of AlexNet, InceptionV3, and ResNet101 were 75%, 81%, and 86%, respectively. Cross-institutional inconsistency and the contributions of image color and subcellular components were studied and analyzed.Gastrointestinal stromal tumors (GIST) are common mesenchymal tumors, and their effective treatment depends upon the mutational subtype of the KIT/PDGFRA genes. We established deep convolutional neural network (DCNN) models to rapidly predict drug-sensitive mutation subtypes from images of pathological tissue. A total of 5153 pathological images of 365 different GISTs from three different laboratories were collected and divided into training and validation sets. A transfer learning mechanism based on DCNN was used with four different network architectures, to identify cases with drug-sensitive mutations. The accuracy ranged from 87% to 75%. Cross-institutional inconsistency, however, was observed. Using gray-scale images resulted in a 7% drop in accuracy (accuracy 80%, sensitivity 87%, specificity 73%). Using images containing only nuclei (accuracy 81%, sensitivity 87%, specificity 73%) or cytoplasm (accuracy 79%, sensitivity 88%, specificity 67%) produced 6% and 8% drops in accuracy rate, respectively, suggesting buffering effects across subcellular components in DCNN interpretation. The proposed DCNN model successfully inferred cases with drug-sensitive mutations with high accuracy. The contribution of image color and subcellular components was also revealed. These results will help to generate a cheaper and quicker screening method for tumor gene testing.

Integer-Only CNNs with 4 Bit Weights and Bit-Shift Quantization Scales at Full-Precision Accuracy

Quantization of neural networks has been one of the most popular techniques to compress models for embedded (IoT) hardware platforms with highly constrained latency, storage, memory-bandwidth, and energy specifications. Limiting the number of bits per weight and activation has been the main focus in the literature. To avoid major degradation of accuracy, common quantization methods introduce additional scale factors to adapt the quantized values to the diverse data ranges, present in full-precision (floating-point) neural networks. These scales are usually kept in high precision, requiring the target compute engine to support a few high-precision multiplications, which is not desirable due to the larger hardware cost. Little effort has yet been invested in trying to avoid high-precision multipliers altogether, especially in combination with 4 bit weights. This work proposes a new quantization scheme, based on power-of-two quantization scales, that works on-par compared to uniform per-channel quantization with full-precision 32 bit quantization scales when using only 4 bit weights. This is done through the addition of a low-precision lookup-table that translates stored 4 bit weights into nonuniformly distributed 8 bit weights for internal computation. All our quantized ImageNet CNNs achieved or even exceeded the Top-1 accuracy of their full-precision counterparts, with ResNet18 exceeding its full-precision model by 0.35%. Our MobileNetV2 model achieved state-of-the-art performance with only a slight drop in accuracy of 0.51%.

On the spatiotemporal generalization of machine learning and ensemble models for simulating built‐up land expansion

AbstractThis study evaluates the spatiotemporal generalization of statistical and machine learning models to simulate built‐up land expansion and compare it to ensemble approaches. Integrated with cellular automata, six individual models—artificial neural networks, support vector machines (SVM), random forest (RF), boosted regression trees, the generalized additive model, the lasso, and two ensemble approaches called ensemble median and ensemble weighted area under the curve—were implemented. Each model was calibrated based on data from 1975–1990, and their extrapolation power was evaluated for 1990–1996, 1996–2000, 2000–2011, and 2011–2017. Total operating characteristics revealed that the RF model achieved the highest calibration accuracy and the highest performance loss during the validation period. The lowest calibration accuracy was related to the SVM model, yet its performance during the validation period increased. In the third time interval (1996–2002), the highest accuracy was again related to the SVM model. A sharp drop in simulation accuracy was seen in all models during the fourth (2002–2011) and fifth intervals (2011–2017). None of the ensemble models appeared to be superior to the individual models. Further, the accuracy of built‐up land expansion models drops noticeably for long‐term simulations.

BS4NN: Binarized Spiking Neural Networks with Temporal Coding and Learning

We recently proposed the S4NN algorithm, essentially an adaptation of backpropagation to multilayer spiking neural networks that use simple non-leaky integrate-and-fire neurons and a form of temporal coding known as time-to-first-spike coding. With this coding scheme, neurons fire at most once per stimulus, but the firing order carries information. Here, we introduce BS4NN, a modification of S4NN in which the synaptic weights are constrained to be binary (+1 or -1), in order to decrease memory (ideally, one bit per synapse) and computation footprints. This was done using two sets of weights: firstly, real-valued weights, updated by gradient descent, and used in the backward pass of backpropagation, and secondly, their signs, used in the forward pass. Similar strategies have been used to train (non-spiking) binarized neural networks. The main difference is that BS4NN operates in the time domain: spikes are propagated sequentially, and different neurons may reach their threshold at different times, which increases computational power. We validated BS4NN on two popular benchmarks, MNIST and Fashion-MNIST, and obtained reasonable accuracies for this sort of network (97.0% and 87.3% respectively) with a negligible accuracy drop with respect to real-valued weights (0.4% and 0.7%, respectively). We also demonstrated that BS4NN outperforms a simple BNN with the same architectures on those two datasets (by 0.2% and 0.9% respectively), presumably because it leverages the temporal dimension. The source codes of the proposed BS4NN are publicly available at https://github.com/SRKH/BS4NN.

DP-Nets: Dynamic programming assisted quantization schemes for DNN compression and acceleration

In this work, we present effective quantization schemes called DP-Nets for the compression and acceleration of deep neural networks (DNNs). A key ingredient is a novel dynamic programming (DP) based algorithm to obtain the optimal solution of scalar K-means clustering. Based on the approaches with regularization and quantization function, two weight quantization approaches called DPR and DPQ for compressing normal DNNs are proposed respectively. Accordingly, a technique based on DP-Nets for inference acceleration is presented. Experiments show that DP-Nets produce models with higher inference accuracy than recently proposed counterparts while achieving same or larger compression. They are also extended for compressing robust DNNs, and the relevant experiments show 16X compression of the robust ResNet-18 model with less than 3% accuracy drop on both natural and adversarial examples. The experiments with FPGA demonstrate that the technique for inference acceleration brings over 5X speedup on matrix–vector multiplication.

The impact of reducing signal acquisition specifications on neuronal spike sorting.

Measuring electrical potentials in the extracellular space of the brain is a popular technique because it can detect action potentials from putative individual neurons. Electrophysiology is undergoing a transformation where the number of recording channels, and thus number of neurons detected, is growing at a dramatic rate. This rapid scaling is paving the way for both new discoveries and commercial applications; however, as the number of channels increases there will be an increasing need to make these systems more power efficient. One area ripe for optimization are the signal acquisition specifications needed to detect and sort action potentials (i.e., "spikes") to putative single neuron sources. In this work, we take existing recordings collected using Intan hardware and modify them in a way that corresponds to reduced recording performance. The accuracy of these degraded recordings to spike sort using MountainSort4 is evaluated by comparing against expert labels. We show that despite reducing signal specifications by a factor of 2 or more, spike sorting accuracy does not change substantially. Specifically, reducing both sample rate and bit depth from 30 kHz and 16 bits to 12 kHz and 12 bits resulted in a 3% drop in spike sorting accuracy. Our results suggest that current neural acquisition systems are over-specified. These results may inform the design of next generation neural acquisition systems enabling higher channel count systems.

Investigation on Robustness of EEG-based Brain-Computer Interfaces.

Electroencephalogram (EEG)-based brain-computer interface (BCI) systems tend to suffer from performance degradation due to the presence of noise and artifacts in EEG data. This study is aimed at systematically investigating the robustness of state-of-the-art machine learning and deep learning based EEG-BCI models for motor imagery classification against simulated channel-specific noise in EEG data, at various low values of signal-to-noise ratio (SNR). Our results illustrate higher robustness of deep learning based MI classification models compared to the traditional machine learning based model, while identifying a set of channels with large sensitivity to simulated channel-specific noise. The EEGNet is relatively more robust towards channel-specific noise than Shallow ConvNet and FBCSP. We propose a preliminary solution, based on activation function, to improve the robustness of the deep learning models. By using saturating nonlinearities, the percentage drop in classification accuracy for SNR of -18 dB had reduced from 10.99% to 6.53% for EEGNet and 14.05% to 3.57% for Shallow ConvNet. Through this study, we emphasize the need for a more precise solution for enhancing the robustness, and thereby usability of EEG-BCI systems.

Channeled imaging spectropolarimeter reconstruction by neural networks.

Snapshot channeled imaging spectropolarimetry (SCISP), which can achieve spectral and polarization imaging without scanning (a single exposure), is a promising optical technique. As Fourier transform is used to reconstruct information, SCISP has its inherent limitations such as channel crosstalk, resolution and accuracy drop, the complex phase calibration, et al. To overcome these drawbacks, a nonlinear technique based on neural networks (NNs) is introduced to replace the role of Fourier reconstruction. Herein, abundant spectral and polarization datasets were built through specially designed generators. The established NNs can effectively learn the forward conversion procedure through minimizing a loss function, subsequently enabling a stable output containing spectral, polarization, and spatial information. The utility and reliability of the proposed technique is confirmed by experiments, which are proved to maintain high spectral and polarization accuracy.

Application of projection methods to simulating mass transport in reverse osmosis systems

Reverse Osmosis has important applications to seawater desalination and advanced water treatment. Its efficiency depends, however, on unsteady fluid flow and solute transport that are challenging to simulate. The challenges arise due to interactions between solute boundary layers and unsteady vortical flow structures generated by complicated geometries. These flow structures also interact with semi-permeable membranes through which the permeate flow depends on the local pressure. We show that this additional pressure coupling causes the temporal accuracy of traditional projection methods to drop to first-order. We track the source of this accuracy drop to the treatment of viscous terms in the derivation of the Poisson equation used to update the velocity and pressure fields. This allows us to propose a modified projection method that recovers second-order temporal accuracy. Finally, we show that the modified projection method can be coupled to convective outlet conditions and immersed boundary conditions to simulate reverse osmosis in steady and unsteady flow regimes.

Empirical observation of negligible fairness–accuracy trade-offs in machine learning for public policy

Growing use of machine learning in policy and social impact settings have raised concerns for fairness implications, especially for racial minorities. These concerns have generated considerable interest among machine learning and artificial intelligence researchers, who have developed new methods and established theoretical bounds for improving fairness, focusing on the source data, regularization and model training, or post-hoc adjustments to model scores. However, little work has studied the practical trade-offs between fairness and accuracy in real-world settings to understand how these bounds and methods translate into policy choices and impact on society. Our empirical study fills this gap by investigating the impact of mitigating disparities on accuracy, focusing on the common context of using machine learning to inform benefit allocation in resource-constrained programs across education, mental health, criminal justice, and housing safety. Here we describe applied work in which we find fairness-accuracy trade-offs to be negligible in practice. In each setting studied, explicitly focusing on achieving equity and using our proposed post-hoc disparity mitigation methods, fairness was substantially improved without sacrificing accuracy. This observation was robust across policy contexts studied, scale of resources available for intervention, time, and relative size of the protected groups. These empirical results challenge a commonly held assumption that reducing disparities either requires accepting an appreciable drop in accuracy or the development of novel, complex methods, making reducing disparities in these applications more practical.

±CIM SRAM for Signed In-Memory Broad-Purpose Computing From DSP to Neural Processing

This work introduces the ±CIM SRAM macro having the unique capability of performing in-memory multiply-and-accumulate computation with signed inputs and signed weights. This uniquely enables the execution of a broad set of workloads, ranging from storage, subsequent signal processing, and pre-conditioning or feature extraction to final convolutional neural network (CNN) computations. The ability to handle arbitrary input/weight sign in any operand within the same array and the same access cycle enables true end-to-end data locality, preserving the inherent benefits of in-memory computing along the entire signal chain. The proposed broad-purpose computing SRAM is based on a commercial 8T dual-port bitcell, and its simplicity allows the enhanced periphery to be pitch-matched with the array, making it amenable for automated design via memory compilers. The ±CIM pipelined architecture allows concurrent read/write and compute operations, avoiding the traditional memory unavailability in compute mode for improved throughput and easier system integration. A 40-nm test chip demonstrating the ±CIM architecture with adjustable input/weight precision exhibits an energy efficiency up to 41 TOPS/W, at an area (energy) overhead of 38% (25%) and negligible performance overhead compared to a compiled SRAM baseline. The sub-LSB computation mean-squared error associated with mismatch (0.38 LSB) and temporal noise (0.62 LSB) confirms the inherent robustness of the architecture. When used for neural network tasks (LeNet-5 and VGG), the accuracy drop is kept between 0.3% and 3.4%, compared to a double-precision software implementation. As an example of digital signal processing (DSP) workload, a frequency-domain feature extractor for voice activity detection keeps the accuracy drop lower than 3.8%.