Bimodal Feature Analysis with Deep Learning for Autism Spectrum Disorder Detection.

Autism Spectrum Disorder (ASD) is a complex and heterogeneous neurodevelopmental disorder which affects a significant proportion of the population, with estimates suggesting that about 1 in 100 children worldwide are affected by ASD. This study introduces a new Deep Neural Network for identifying ASD in children through gait analysis, using features extracted from frames composing video recordings of their walking patterns. The innovative method presented herein is based on imagery and combines gait analysis and deep learning, offering a noninvasive and objective assessment of neurodevelopmental disorders while delivering high accuracy in ASD detection. Our model proposes a bimodal approach based on the concatenation of two distinct Convolutional Neural Networks processing two feature sets extracted from the same videos. The features obtained from the convolutions of both networks are subsequently flattened and merged into a single vector, serving as input for the fully connected layers in the binary classification process. This approach demonstrates the potential for effective ASD detection in children through the combination of gait analysis and deep learning techniques.

An adaptive embedding procedure for time series forecasting with deep neural networks.

Nowadays, solving time series prediction problems is an open and challenging task. Many solutions are based on the implementation of deep neural architectures, which are able to analyze the structure of the time series and to carry out the prediction. In this work, we present a novel deep learning scheme based on an adaptive embedding mechanism. The latter is exploited to extract a compressed representation of the input time series that is used for the subsequent forecasting. The proposed model is based on a two-layer bidirectional Long Short-Term Memory network, where the first layer performs the adaptive embedding and the second layer acts as a predictor. The performances of the proposed forecasting scheme are compared with several models in two different scenarios, considering both well-known time series and real-life application cases. The experimental results show the accuracy and the flexibility of the proposed approach, which can be used as a prediction tool for any actual application.

Multi-Damage Detection in Composite Space Structures via Deep Learning.

The diagnostics of environmentally induced damages in composite structures plays a critical role for ensuring the operational safety of space platforms. Recently, spacecraft have been equipped with lightweight and very large substructures, such as antennas and solar panels, to meet the performance demands of modern payloads and scientific instruments. Due to their large surface, these components are more susceptible to impacts from orbital debris compared to other satellite locations. However, the detection of debris-induced damages still proves challenging in large structures due to minimal alterations in the spacecraft global dynamics and calls for advanced structural health monitoring solutions. To address this issue, a data-driven methodology using Long Short-Term Memory (LSTM) networks is applied here to the case of damaged solar arrays. Finite element models of the solar panels are used to reproduce damage locations, which are selected based on the most critical risk areas in the structures. The modal parameters of the healthy and damaged arrays are extracted to build the governing equations of the flexible spacecraft. Standard attitude manoeuvres are simulated to generate two datasets, one including local accelerations and the other consisting of piezoelectric voltages, both measured in specific locations of the structure. The LSTM architecture is then trained by associating each sensed time series with the corresponding damage label. The performance of the deep learning approach is assessed, and a comparison is presented between the accuracy of the two distinct sets of sensors: accelerometers and piezoelectric patches. In both cases, the framework proved effective in promptly identifying the location of damaged elements within limited measured time samples.

Perceptron Theory Can Predict the Accuracy of Neural Networks.

Multilayer neural networks set the current state of the art for many technical classification problems. But, these networks are still, essentially, black boxes in terms of analyzing them and predicting their performance. Here, we develop a statistical theory for the one-layer perceptron and show that it can predict performances of a surprisingly large variety of neural networks with different architectures. A general theory of classification with perceptrons is developed by generalizing an existing theory for analyzing reservoir computing models and connectionist models for symbolic reasoning known as vector symbolic architectures. Our statistical theory offers three formulas leveraging the signal statistics with increasing detail. The formulas are analytically intractable, but can be evaluated numerically. The description level that captures maximum details requires stochastic sampling methods. Depending on the network model, the simpler formulas already yield high prediction accuracy. The quality of the theory predictions is assessed in three experimental settings, a memorization task for echo state networks (ESNs) from reservoir computing literature, a collection of classification datasets for shallow randomly connected networks, and the ImageNet dataset for deep convolutional neural networks. We find that the second description level of the perceptron theory can predict the performance of types of ESNs, which could not be described previously. Furthermore, the theory can predict deep multilayer neural networks by being applied to their output layer. While other methods for prediction of neural networks performance commonly require to train an estimator model, the proposed theory requires only the first two moments of the distribution of the postsynaptic sums in the output neurons. Moreover, the perceptron theory compares favorably to other methods that do not rely on training an estimator model.

A Study on Structural Health Monitoring of a Large Space Antenna via Distributed Sensors and Deep Learning.

Most modern Earth and Universe observation spacecraft are now equipped with large lightweight and flexible structures, such as antennas, telescopes, and extendable elements. The trend of hosting more complex and bigger appendages, essential for high-precision scientific applications, made orbiting satellites more susceptible to performance loss or degradation due to structural damages. In this scenario, Structural Health Monitoring strategies can be used to evaluate the health status of satellite substructures. However, in particular when analysing large appendages, traditional approaches may not be sufficient to identify local damages, as they will generally induce less observable changes in the system dynamics yet cause a relevant loss of payload data and information. This paper proposes a deep neural network to detect failures and investigate sensor sensitivity to damage classification for an orbiting satellite hosting a distributed network of accelerometers on a large mesh reflector antenna. The sensors-acquired time series are generated by using a fully coupled 3D simulator of the in-orbit attitude behaviour of a flexible satellite, whose appendages are modelled by using finite element techniques. The machine learning architecture is then trained and tested by using the sensors' responses gathered in a composite scenario, including not only the complete failure of a structural element (structural break) but also an intermediate level of structural damage. The proposed deep learning framework and sensors configuration proved to accurately detect failures in the most critical area or the structure while opening new investigation possibilities regarding geometrical properties and sensor distribution.

A comparison of machine learning classifiers for smartphone-based gait analysis.

This paper proposes a reliable monitoring scheme that can assist medical specialists in watching over the patient's condition. Although several technologies are traditionally used to acquire motion data of patients, the high costs as well as the large spaces they require make them difficult to be applied in a home context for rehabilitation. A reliable patient monitoring technique, which can automatically record and classify patient movements, is mandatory for a telemedicine protocol. In this paper, a comparison of several state-of-the-art machine learning classifiers is proposed, where stride data are collected by using a smartphone. The main goal is to identify a robust methodology able to assure a suited classification of gait movements, in order to allow the monitoring of patients in time as well as to discriminate among a pathological and physiological gait. Additionally, the advantages of smartphones of being compact, cost-effective and relatively easy to operate make these devices particularly suited for home-based rehabilitation programs. Graphical Abstract. This paper proposes a reliable monitoring scheme that can assist medical specialists in watching over the patient's condition. Although several technologies are traditionally used to acquire motion data of patients, the high costs as well as the large spaces they require make them difficult to be applied in a home context for rehabilitation. A reliable patient monitoring technique, which can automatically record and classify patient movements, is mandatory for a telemedicine protocol. In this paper, a comparison of several state-of-the-art machine learning classifiers is proposed, where stride data are collected and processed by using a smartphone(see figure). The main goal is to identify a robust methodology able to assure a suited classification of gait movements, in order to allow the monitoring of patients in time as well as to discriminate among a pathological and physiological gait. Additionally, the advantages of smartphones of being compact, cost-effective and relatively easy to operate make these devices particularly suited for home-based rehabilitation programs.

Power spectra prognostic aspects of impulsive eye movement traces in superior vestibular neuritis.

Since usefulness of power spectra (PS) analysis was demonstrated in many fields of electrophysiology, the aims of the present study were to evaluate differences in frequency domain values of eye velocity traces between a group of 60 healthy subjects (HC) and 35 matched superior vestibular neuritis (VN) patients and to determine prognostic aspects of such values in terms of superior VN recovery. PS calculated on video head impulse test traces was compared between HC and during the acute stage of vertigo (T1) and after 3 months (T2) in superior VN patients. A multiple regression and desirability model between Δ (T2gain - T1gain) vestibulo-ocular reflex (VOR) gain and five prognostic factors were employed. Significant PS differences within the 7.8-16.6 Hz domain were found between superior VN and HC. A significant negative correlation was found between the 7.8-16.6 Hz domain unitary PS value and Δ VOR gain (ß = - 0.836). The desirability model depicted a cutoff value of the unitary PS equal to 1.82 in order to obtain a Δ VOR gain rate at least equal to 0.1. Present findings could be a further step for monitoring those superior VN patients with systemic risk factors and high risk of VOR incomplete recovery. Graphical abstract In the top left, healthy control (HC) and superior vestibular neuritis (VN) subjects were screened by means of video head impulse test. In the top right, significant differences in power spectra values were depicted within the 7.8-16.6 Hz domain when comparing the two groups of subjects. In the bottom center, the desirability model depicts a cutoff value of the power spectra equal to 1.82 in order to obtain a Δ vestibulo-ocular reflex gain rate at least equal to 0.1.

Fully Decentralized Semi-supervised Learning via Privacy-preserving Matrix Completion.

Distributed learning refers to the problem of inferring a function when the training data are distributed among different nodes. While significant work has been done in the contexts of supervised and unsupervised learning, the intermediate case of Semi-supervised learning in the distributed setting has received less attention. In this paper, we propose an algorithm for this class of problems, by extending the framework of manifold regularization. The main component of the proposed algorithm consists of a fully distributed computation of the adjacency matrix of the training patterns. To this end, we propose a novel algorithm for low-rank distributed matrix completion, based on the framework of diffusion adaptation. Overall, the distributed Semi-supervised algorithm is efficient and scalable, and it can preserve privacy by the inclusion of flexible privacy-preserving mechanisms for similarity computation. The experimental results and comparison on a wide range of standard Semi-supervised benchmarks validate our proposal.

A novel fuzzy approach for automatic Brunnstrom stage classification using surface electromyography.

Clinical assessment plays a major role in post-stroke rehabilitation programs for evaluating impairment level and tracking recovery progress. Conventionally, this process is manually performed by clinicians using chart-based ordinal scales which can be both subjective and inefficient. In this paper, a novel approach based on fuzzy logic is proposed which automatically evaluates stroke patients' impairment level using single-channel surface electromyography (sEMG) signals and generates objective classification results based on the widely used Brunnstrom stages of recovery. The correlation between stroke-induced motor impairment and sEMG features on both time and frequency domain is investigated, and a specifically designed fuzzy kernel classifier based on geometrically unconstrained membership function is introduced in the study to tackle the challenges in discriminating data classes with complex separating surfaces. Experiments using sEMG data collected from stroke patients have been carried out to examine the validity and feasibility of the proposed method. In order to ensure the generalization capability of the classifier, a cross-validation test has been performed. The results, verified using the evaluation decisions provided by an expert panel, have reached a rate of success of the 92.47%. The proposed fuzzy classifier is also compared with other pattern recognition techniques to demonstrate its superior performance in this application.

Selection of clinical features for pattern recognition applied to gait analysis.

This paper deals with the opportunity of extracting useful information from medical data retrieved directly from a stereophotogrammetric system applied to gait analysis. A feature selection method to exhaustively evaluate all the possible combinations of the gait parameters is presented, in order to find the best subset able to classify among diseased and healthy subjects. This procedure will be used for estimating the performance of widely used classification algorithms, whose performance has been ascertained in many real-world problems with respect to well-known classification benchmarks, both in terms of number of selected features and classification accuracy. Precisely, support vector machine, Naive Bayes and K nearest neighbor classifiers can obtain the lowest classification error, with an accuracy greater than 97 %. For the considered classification problem, the whole set of features will be proved to be redundant and it can be significantly pruned. Namely, groups of 3 or 5 features only are able to preserve high accuracy when the aim is to check the anomaly of a gait. The step length and the swing speed are the most informative features for the gait analysis, but also cadence and stride may add useful information for the movement evaluation.

Distributed semi-supervised support vector machines.

The semi-supervised support vector machine (S(3)VM) is a well-known algorithm for performing semi-supervised inference under the large margin principle. In this paper, we are interested in the problem of training a S(3)VM when the labeled and unlabeled samples are distributed over a network of interconnected agents. In particular, the aim is to design a distributed training protocol over networks, where communication is restricted only to neighboring agents and no coordinating authority is present. Using a standard relaxation of the original S(3)VM, we formulate the training problem as the distributed minimization of a non-convex social cost function. To find a (stationary) solution in a distributed manner, we employ two different strategies: (i) a distributed gradient descent algorithm; (ii) a recently developed framework for In-Network Nonconvex Optimization (NEXT), which is based on successive convexifications of the original problem, interleaved by state diffusion steps. Our experimental results show that the proposed distributed algorithms have comparable performance with respect to a centralized implementation, while highlighting the pros and cons of the proposed solutions. To the date, this is the first work that paves the way toward the broad field of distributed semi-supervised learning over networks.

A decentralized training algorithm for Echo State Networks in distributed big data applications.

The current big data deluge requires innovative solutions for performing efficient inference on large, heterogeneous amounts of information. Apart from the known challenges deriving from high volume and velocity, real-world big data applications may impose additional technological constraints, including the need for a fully decentralized training architecture. While several alternatives exist for training feed-forward neural networks in such a distributed setting, less attention has been devoted to the case of decentralized training of recurrent neural networks (RNNs). In this paper, we propose such an algorithm for a class of RNNs known as Echo State Networks. The algorithm is based on the well-known Alternating Direction Method of Multipliers optimization procedure. It is formulated only in terms of local exchanges between neighboring agents, without reliance on a coordinating node. Additionally, it does not require the communication of training patterns, which is a crucial component in realistic big data implementations. Experimental results on large scale artificial datasets show that it compares favorably with a fully centralized implementation, in terms of speed, efficiency and generalization accuracy.

A Fuzzy Kernel Motion Classifier for Autonomous Stroke Rehabilitation.

Autonomous poststroke rehabilitation systems which can be deployed outside hospital with no or reduced supervision have attracted increasing amount of research attentions due to the high expenditure associated with the current inpatient stroke rehabilitation systems. To realize an autonomous systems, a reliable patient monitoring technique which can automatically record and classify patient's motion during training sessions is essential. In order to minimize the cost and operational complexity, the combination of nonvisual-based inertia sensing devices and pattern recognition algorithms are often considered more suitable in such applications. However, the high motion irregularity due to stroke patients' body function impairment has significantly increased the classification difficulty. A novel fuzzy kernel motion classifier specifically designed for stroke patient's rehabilitation training motion classification is presented in this paper. The proposed classifier utilizes geometrically unconstrained fuzzy membership functions to address the motion class overlapping issue, and thus, it can achieve highly accurate motion classification even with poorly performed motion samples. In order to validate the performance of the classifier, experiments have been conducted using real motion data sampled from stroke patients with a wide range of impairment level and the results have demonstrated that the proposed classifier is superior in terms of error rate compared to other popular algorithms.