Complexity-accuracy Trade-off Research Articles

Comparing machine-learning models of different levels of complexity for crop protection: A look into the complexity-accuracy tradeoff

Crop diseases and pests constitute significant causes of yield losses for crops. To limit the harm incurred by those events, farmers resort to plant protection products. Such products are known to have adverse effects both on the environment and on human health. Agronomists make continuous efforts to limit the usage of plant protection products to situations where those products are strictly necessary. To determine such situations, agronomists and policy-makers often rely on decision support tools to model and predict the dynamics of plant diseases. Decision support tools are based either on mechanistic models or on statistical approaches learned from large datasets of biotic (e.g., disease incidence, plant phenological stage) and abiotic (meteorological, soil characteristics) observations in cultures. The surge of powerful machine learning (ML) methods in the last decade makes such approaches a natural pathway to model the dynamics of plant diseases.Machine learning models can reveal the factors that contribute the most to disease and pests outbreaks, provided that those models are simple enough for human inspection. Simplicity, however, may come at the price of lower prediction performances when compared to more complex models.In this paper, we offer a deep look at the performance of ML models of different complexity when used on two use cases of crop disease prediction: downy mildew in the grapevine, and Cercospora leaf spot in the sugar beet.We compare model accuracy and complexity using a year-based cross-validation approach. Our results suggest that interannual meteorological variations are a very important factor in plant disease prediction. Moreover, in line with the observations of the research community in interpretable ML, model complexity stands in clear trade-off with accuracy. This makes models of intermediate complexity appealing for predicting the dynamics of crop diseases as they can provide explicit insights about the rationale of their predictions.

Beam SNR Prediction Using Channel Charting

We consider a machine learning approach for beam handover in mmWave 5G New Radio systems, in which User Equipments (UEs) perform autonomous beam selection, conditioned on a used Base Station (BS) beam. We develop a network-centric approach for predicting beam Signal-to-Noise Ratio (SNR) from Channel State Information (CSI) features measured at the BS, which consists of two phases; offline and online. In the offline training phase, we construct CSI features and dimensionality-reduced Channel Charts (CC). We annotate the CCs with per-beam SNRs for different combinations of a BS beam and the corresponding best UE beam, and train models to predict SNR from CSI features for different BS/UE beam combinations. In the online phase, we predict SNRs of beam combinations not being used at the moment. We develop a low complexity out-of-sample algorithm for dimensionality reduction in the online phase. We consider <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$k$</tex-math></inline-formula> -nearest neighbors, Gaussian process regression, and neural network-based predictions. To evaluate the efficacy of the proposed framework, we perform simulations for a street segment with synthetically generated CSI. We investigate the complexity-accuracy trade-off for different dimensionality reduction techniques and different predictors. Our results reveal that nonlinear dimensionality reduction of CSI features with neural network prediction shows the best performance, and the performance of the best CSI-based prediction method is comparable to prediction based on using known physical location.

The trade-off between complexity and accuracy. Preparing for computer-based adaptive instruction on fractions

ABSTRACT PS-I learning approaches with an initial problem-solving phase (PS) followed by an instruction phase (I) can be optimised by providing adaptive (individualised) instruction that builds on students’ problem-solving products. A computer-based implementation of both phases offers the possibility of adaptivity at the transition from the problem-solving to the instruction phase. However, it also poses two challenges for the problem-solving phase, which should allow for (1) complex problem-solving processes (process validity) and (2) accurate diagnosis of problem-solving products. Increased openness of the learning environment can increase the process validity and decrease diagnostic accuracy and vice-versa (complexity-accuracy trade-off). Using two variants of a computer-based learning environment for fraction comparison with different extent of openness (O+ and O-), we investigate whether both requirements (complex problem-solving processes and accurate diagnosis of the problem-solving products) can be met simultaneously. To check process validity (complexity of problem-solving processes), we compare the problem-solving products with those from previous research on an analogous paper-based version. Our results show that valid and comparable problem-solving processes can be achieved in both computer-based learning environments (O+ and O-). Moreover, only the learning environment O- enables an accurate computer-based interpretation of students’ products and thus resolving the complexity-accuracy trade-off.

RFDOA-Net: An Efficient ConvNet for RF-Based DOA Estimation in UAV Surveillance Systems

This paper presents a convolution neural network (CNN)-based direction of arrival (DOA) estimation method for radio frequency (RF) signals acquired by a nonuniform linear antenna array (NULA) in unmanned aerial vehicle (UAV) localization systems. The proposed deep CNN, namely RFDOA-Net, is designed with three primary processing modules, such as collective feature extraction, multi-scaling feature processing, and complexity-accuracy trade-off, to learn the multi-scale intrinsic characteristics for multi-class angle classification. In several specific modules, the regular convolutional and grouped convolutional layers are leveraged with different filter sizes to enrich diversified features and reduce network complexity besides adopting residual connection to prevent vanishing gradient. For performance evaluation, we generate a synthetic signal dataset for DOA estimation under the multipath propagation channel with the presence of additive noise, propagation attenuation and delay. In simulations, the effectiveness of RFDOA-Net is investigated comprehensively with various processing modules and antenna configurations. Compared with several state-of-the-art deep learning-based models, RFDOA-Net shows the superiority in terms of accuracy with over 94% accuracy at 5 dB signal-to-noise ratio (SNR) with cost-efficiency.

Real-Time Parallel-Serial LiDAR-Based Localization Algorithm with Centimeter Accuracy for GPS-Denied Environments.

In this paper, we introduce a real-time parallel-serial algorithm for autonomous robot positioning for GPS-denied, dark environments, such as caves and mine galleries. To achieve a good complexity-accuracy trade-off, we fuse data from light detection and ranging (LiDAR) and an inertial measurement unit (IMU). The proposed algorithm’s main novelty is that, unlike in most algorithms, we apply an extended Kalman filter (EKF) to each LiDAR scan point and calculate the location relative to a triangular mesh. We also introduce three implementations of the algorithm: serial, parallel, and parallel-serial. The first implementation verifies the correctness of our innovative approach, but is too slow for real-time execution. The second approach implements a well-known parallel data fusion approach, but is still too slow for our application. The third and final implementation of the presented algorithm along with the state-of-the-art GPU data structures achieves real-time performance. According to our experimental findings, our algorithm outperforms the reference Gaussian mixture model (GMM) localization algorithm in terms of accuracy by a factor of two.

Double Bayesian Smoothing as Message Passing

Recently, a novel method for developing filtering algorithms, based on the interconnection of two Bayesian filters and called double Bayesian filtering, has been proposed. In this manuscript we show that the same conceptual approach can be exploited to devise a new smoothing method, called double Bayesian smoothing. A double Bayesian smoother combines a double Bayesian filter, employed in its forward pass, with the interconnection of two backward information filters used in its backward pass. As a specific application of our general method, a detailed derivation of double Bayesian smoothing algorithms for conditionally linear Gaussian systems is illustrated. Numerical results for two specific dynamic systems evidence that these algorithms can achieve a better complexity-accuracy tradeoff and tracking capability than other smoothing techniques recently appeared in the literature.

Multiple Bayesian Filtering as Message Passing

In this manuscript, a general method for deriving filtering algorithms that involve a network of interconnected Bayesian filters is proposed. This method is based on the idea that the processing accomplished inside each of the Bayesian filters and the interactions between them can be represented as message passing algorithms over a proper graphical model. The usefulness of our method is exemplified by developing new filtering techniques, based on the interconnection of a particle filter and an extended Kalman filter, for conditionally linear Gaussian systems. Numerical results for two specific dynamic systems evidence that the devised algorithms can achieve a better complexity-accuracy tradeoff than marginalized particle filtering and multiple particle filtering.

Multi-fidelity deep neural networks for adaptive inference in the internet of multimedia things

Internet of Things (IoT) infrastructures are more and more relying on multimedia sensors to provide information about the environment. Deep neural networks (DNNs) could extract knowledge from this audiovisual data but they typically require large amounts of resources (processing power, memory and energy). If all limitations of the execution environment are known beforehand, we can design neural networks under these constraints. An IoT setting however is a very heterogeneous environment where the constraints can change rapidly. We propose a technique allowing us to deploy a variety of different networks at runtime, each with a specific complexity-accuracy trade-off but without having to store each network independently. We train a sequence of networks of increasing size and constrain each network to contain the parameters of all smaller networks in the sequence. We only need to store the largest network to be able to deploy each of the smaller networks. We experimentally validate our approach on different benchmark datasets for image recognition and conclude that we can build networks that support multiple trade-offs between accuracy and computational cost.

Hysteresis nonlinearity modeling and linearization approach for Envelope Tracking Power Amplifiers in wireless systems

Novel Dual-Input Single-Output (DISO) behavioral modeling and linearization techniques for Envelope Tracking Power Amplifiers (ET PAs) in wireless application are presented in this paper. The proposed modeling approach applies the Hammerstein structure for the Amplitude-to-Phase (AM/PM) conversion. An extension of the Saleh AM/PM model is also proposed to model the static nonlinearity in the ET PA AM/PM conversion. This paper presents a new linearization technique for ET PAs by inverting the proposed ET PA model and using it as a Digital Predistortion (DPD). The measurement results show that both the proposed model and DPD can clearly outperform both the original Saleh model and the DPD based on the Saleh model inversion. Compared to the state-of-the-art behavioral models for ET PAs, the proposed model and DPD can offer an improved complexity-accuracy trade-off thanks to the lower number of coefficients.

Maximum-Variance Total Variation Denoising for Interpretable Spatial Smoothing

We consider the problem of spatial regression where interpretability of the model is a high priority. Such problems appear frequently in a diverse set of fields from climatology to epidemiology to predictive policing. For cognitive, logistical, and organizational reasons, humans tend to infer regions or neighborhoods of constant value, often with sharp discontinuities between regions, and then assign resources on a per-region basis. Automating this smoothing process presents a unique challenge for spatial smoothing algorithms, which tend to assume stationarity and smoothness everywhere. To address this problem, we propose Maximum Variance Total Variation (MVTV) denoising, a novel method for interpretable nonlinear spatial regression. MVTV divides the feature space into blocks of constant value and smooths the value of all blocks jointly via a convex optimization routine. Our method is fully data-adaptive and incorporates highly robust routines for tuning all hyperparameters automatically. We compare our approach against the existing CART and CRISP methods via both a complexity-accuracy tradeoff metric and a human study, demonstrating that that MVTV is a more powerful and interpretable method.

A Petri net tool for software performance estimation based on upper throughput bounds

Functional and non-functional properties analysis (i.e., dependability, security, or performance) ensures that requirements are fulfilled during the design phase of software systems. However, the Unified Modelling Language (UML), standard de facto in industry for software systems modelling, is unsuitable for any kind of analysis but can be tailored for specific analysis purposes through profiling. For instance, the MARTE profile enables to annotate performance data within UML models that can be later transformed to formal models (e.g., Petri nets or timed automatas) for performance evaluation. A performance (or throughput) estimation in such models normally relies on a whole exploration of the state space, which becomes unfeasible for large systems. To overcome this issue upper throughput bounds are computed, which provide an approximation to the real system throughput with a good complexity-accuracy trade-off. This paper introduces a tool, named PeabraiN, that estimates the performance of software systems via their UML models. To do so, UML models are transformed to Petri nets where performance is estimated based on upper throughput bounds computation. PeabraiN also allows to compute other features on Petri nets, such as the computation of upper and lower marking place bounds, and to simulate using an approximate (continuous) method. We show the applicability of PeabraiN by evaluating the performance of a building closed circuit TV system.

The dispersal analysis on basis construction of digital predistortion techniques for power amplifiers

Digital predistortion (DPD) exploiting the Volterra series can effectively compensate the strong nonlinearities (with memory effects) of power amplifiers. Diverse pruning schemes have been proposed to reduce the computational complexity with comparative performances. Regrettably, the mechanism of basis construction in different DPDs that is related to the complexity-efficiency tradeoff has not been well-studied. In order to compare methodically the Volterra-series based DPDs, the key parameters of the basis are studied here. Especially for the cross terms that describe the interaction between nonlinearities and memory effect, the complexity-efficiency tradeoff cannot be addressed directly. Thus, the basis construction of DPDs is firstly mapped to the general Volterra-series to observe the pruning property of each DPD. The diagonal dispersal property on basis construction is concluded as the most efficient and reliable way for pruning DPD techniques. The running complexity-accuracy tradeoff is quantitatively analyzed via the number of floating point operations. With the best performance settings found by the Qhull algorithm, the experimental results show that (1) the nonlinear order of the cross terms is more significant than the related memory depth, and (2) basis construction with diagonal dispersal properties effectively shows the complexity-efficiency tradeoff of different DPDs. These results offer new angles to assess the feasibility of DPDs and their involved complexity.

A semi-algebraic framework for approximate CP decompositions via simultaneous matrix diagonalizations (SECSI)

In this paper, we propose a framework to compute approximate CANDECOMP / PARAFAC (CP) decompositions. Such tensor decompositions are viable tools in a broad range of applications, creating the need for versatile tools to compute such decompositions with an adjustable complexity-accuracy trade-off.To this end, we propose a novel SEmi-algebraic framework that allows the computation of approximate C P decompositions via SImultaneous Matrix Diagonalizations (SECSI). In contrast to previous Simultaneous Matrix Diagonalization (SMD)-based approaches, we use the tensor structure to construct not only one but the full set of possible SMDs. Solving all SMDs, we obtain multiple estimates of the factor matrices and present strategies to choose the best estimate in a subsequent step. This SECSI framework retains the option to choose the number of SMDs to solve and to adopt various strategies for the selection of the final solution out of the multiple estimates. A best matching scheme based on an exhaustive search as well as heuristic selection schemes are devised to flexibly adapt to specific applications. Four example algorithms with different accuracy-complexity trade-off points are compared to state-of-the-art algorithms. We obtain more reliable estimates and a reduced computational complexity.

Design of a Low-Power On-Body ECG Classifier for Remote Cardiovascular Monitoring Systems

In this paper, we first present a detailed study on the trade-off between the computational complexity (directly related to the power consumption) and classification accuracy for a number of classifiers for classifying normal and abnormal electrocardiograms (ECGs). In our analysis, we consider the spectral energy of the constituent waves of the ECG as the discriminative feature. Starting with the exhaustive exploration of single heartbeat-based classification to ascertain the complexity-accuracy trade-off in different classification algorithms, we then extend our study for multiple heartbeat-based classification. We use data available in Physionet as well as samples from Southampton General Hospital Cardiology Department for our investigation. Our primary conclusion is that a classifier based on linear discriminant analysis (LDA) achieves comparable level of accuracy to the best performing support vector machine classifiers with advantage of significantly reduced computational complexity. Subsequently, we propose an ultra low-power circuit implementation of the LDA classifier that could be integrated with the ECG sensor node enabling on-body normal and abnormal ECG classification. The simulated circuit is synthesized at 130 nm technology and occupies 0.70 mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of silicon area (0.979 mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> after place and route) while it consumes 182.94 nW @ 1.08 V, estimated with Synopsys PrimeTime when operating at 1 KHz. These results clearly demonstrate the potential for low-power implementation of the proposed design.

A Novel Selective Scale Space based Fuzzy C-means Model for Spatial Clustering

Abstract This paper proposed a novel Scale Space Filter based Fuzzy C-Means algorithm for clustering spatial data. The number of clusters, C, in present case is known in advance. The Scale Space filter is used for better separability of the data which are not linearly separable and in the present paper the same is used to selective parameters for betterment to meet the complexity-accuracy tradeoff. The Xie-Beni validity index is used as Objective Function of the model to check the quality of the clusters produced. The Results are tested on Standard iris data. The analysis and comparative study with existing algorithms has also been drawn.

Reducing the Complexity of the N-FINDR Algorithm for Hyperspectral Image Analysis

The N-FINDR algorithm for unmixing hyperspectral data is both popular and successful. However, opportunities for improving the algorithm exist, particularly to reduce its computational expense. Two approaches to achieve this are examined. First, the redundancy inherent in the determinant calculations at the heart of N-FINDR is reduced using an LDU decomposition to form two new algorithms, one based on the original N-FINDR algorithm and one based on the closely related Sequential N-FINDR algorithm. The second approach lowers complexity by reducing the repetition of the volume calculations by removing pixels unlikely to represent pure materials. This is accomplished at no additional cost through the reuse of the volume calculations inherent in the Sequential N-FINDR algorithm. Various thresholding methods for excluding pixels are considered. The impact of these modifications on complexity and the accuracy is examined on simulated and real data showing that the LDU-based approaches save considerable complexity, while pixel reduction methods, with appropriate threshold selection, can produce a favorable complexity-accuracy trade-off.

A* orthogonal matching pursuit: Best-first search for compressed sensing signal recovery

Compressed sensing is a developing field aiming at reconstruction of sparse signals acquired in reduced dimensions, which make the recovery process under-determined. The required solution is the one with minimum $\ell_0$ norm due to sparsity, however it is not practical to solve the $\ell_0$ minimization problem. Commonly used techniques include $\ell_1$ minimization, such as Basis Pursuit (BP) and greedy pursuit algorithms such as Orthogonal Matching Pursuit (OMP) and Subspace Pursuit (SP). This manuscript proposes a novel semi-greedy recovery approach, namely A* Orthogonal Matching Pursuit (A*OMP). A*OMP performs A* search to look for the sparsest solution on a tree whose paths grow similar to the Orthogonal Matching Pursuit (OMP) algorithm. Paths on the tree are evaluated according to a cost function, which should compensate for different path lengths. For this purpose, three different auxiliary structures are defined, including novel dynamic ones. A*OMP also incorporates pruning techniques which enable practical applications of the algorithm. Moreover, the adjustable search parameters provide means for a complexity-accuracy trade-off. We demonstrate the reconstruction ability of the proposed scheme on both synthetically generated data and images using Gaussian and Bernoulli observation matrices, where A*OMP yields less reconstruction error and higher exact recovery frequency than BP, OMP and SP. Results also indicate that novel dynamic cost functions provide improved results as compared to a conventional choice.

CATO: complexity-accuracy trade-off localisation approach in wireless sensor networks

A low computational complexity localisation algorithm with an innovative test procedure and an approximate maximum likelihood-least square hybrid estimation scheme is proposed. The work is based on the following principles: the residual squares from line of sight propagation follow a central chi-square distribution of one degree of freedom; a portion of the residual squares follow a central chi-square distribution, if all the normalised residual squares follow the same distribution. The measures can approximately reduce computational complexity by 90%.

Substate tying with combined parameter training and reduction in tied-mixture HMM design

Two approaches are proposed for the design of tied-mixture hidden Markov models (TMHMM). One approach improves parameter sharing via partial tying of TMHMM states. To facilitate tying at the substate level, the state emission probabilities are constructed in two stages or, equivalently, are viewed as a mixture of mixtures of Gaussians. This paradigm allows, and is complemented with, an optimization technique to seek the best complexity-accuracy tradeoff solution, which jointly exploits Gaussian density sharing and substate tying. Another approach to enhance model training is combined training and reduction of model parameters. The procedure starts by training a system with a large universal codebook of Gaussian densities. It then iteratively reduces the size of both the codebook and the mixing coefficient matrix, followed by parameter re-training. The additional cost in design complexity is modest. Experimental results on the ISOLET database and its E-set subset show that substate tying reduces the classification error rate by over 15%, compared to standard Gaussian sharing and whole-state tying. TMHMM design with combined training and reduction of parameters reduces the classification error rate by over 20% compared to conventional TMHMM design. When the two proposed approaches were integrated, 25% error rate reduction over TMHMM with whole-state tying was achieved.